Buckets:
| {"schema_version":1,"generated_at":"2026-07-25T17:09:11Z","docs":[{"root_id":"2081060081278558271","url":"https://x.com/sudoingX/status/2081060081278558271","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T16:52:42.000Z","lang":"en","categories":["context","hardware","models","performance"],"models":[],"posts":[{"id":"2081060081278558271","text":"a few days ago i benchmarked the mechanics of bonsai 27b on an 8gb gaming card. the speed, the window, the footprint, all measured single stream on a $200 rtx 3060 ti.\n\nfor anyone new here, bonsai is a 1bit crush of qwen 3.6 27b, the model i spent all month calling the king of the 3090, squeezed to 3.9gb. small enough to live on a card from a budget gaming rig.\n\nthe numbers held. 42 tok/s, 2x what the 6gb card manages. the full server loads into 6.8 of 8 gigs at a 128k context, twice what an agent actually needs. the 8gb tier stopped being a chatbot tier and quietly became a 128k agent tier.\n\nbut speed is the easy half. the real question is whether it can actually WORK. so now i put it to the test.\n\nthe challenge now is one prompt, zero steering, on the 8gb card: build a working wankel rotary engine from scratch. the epitrochoid housing, the rotor turning at exactly one third of shaft speed, three combustion chambers cycling intake, compression, power, exhaust. and it has to pass a suite of deterministic geometry tests to count as done. not a drawing, a provably correct simulation.\n\ni spun up a small channel for these builds, the run goes live there, and i'm opensourcing the whole benchmark suite after.\n\na 1bit model, on a card that costs less than a phone, building a working engine from one sentence. this is the part i've been waiting to show you. exact flags and the full model card in the reply.","created_at":"2026-07-25T16:52:42.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOFiu_QbYAAdavu.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":2,"views":591},"quoted_tweet_id":"2079180442285244814"}]},{"root_id":"2081044126917611992","url":"https://x.com/StefanMaier/status/2081044126917611992","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-25T15:49:18.000Z","lang":"en","categories":["local","models"],"models":[],"posts":[{"id":"2081044126917611992","text":"Resetting my local LLM (brain).\nSince the biggest folk festival in my hometown has kicked off, I’m swapping LLMs for beer with \"large parameters.\" https://t.co/Qhc0maDVIF","created_at":"2026-07-25T15:49:18.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOFamf8XMAAHFK2.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HOFamhkWMAAettP.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":0,"views":27},"quoted_tweet_id":null}]},{"root_id":"2081031099522674696","url":"https://x.com/LyalinDotCom/status/2081031099522674696","author":{"username":"LyalinDotCom","display_name":"Dmitry Lyalin"},"created_at":"2026-07-25T14:57:32.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2081031099522674696","text":"I don’t think people always know just how much Google is pro open source. And I mean look at all we’ve done with Gemma.","created_at":"2026-07-25T14:57:32.000Z","media":[],"metrics":{"likes":239,"retweets":6,"replies":26,"views":8596},"quoted_tweet_id":"2081026488158040181"}]},{"root_id":"2081001466261811522","url":"https://x.com/LottoLabs/status/2081001466261811522","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-25T12:59:47.000Z","lang":"en","categories":["local","performance"],"models":[],"posts":[{"id":"2081001466261811522","text":"Building a bunch of local first terminal bench style benchmarks designed for local models first\n\nTB2.1 are great and they are proxies for how well local models will perform on a bunch of tasks but they’re designed for sota models\n\nThese benchmarks for local models will be simplified in design and more relevant to smaller models and how well they perform at typical tasks I’ve seen users run as assistants and research bots, there will still be coding tasks but they are designed to reflect actual realistic use cases, including partial scoring","created_at":"2026-07-25T12:59:47.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":84},"quoted_tweet_id":null}]},{"root_id":"2080992214499336461","url":"https://x.com/TheZachMueller/status/2080992214499336461","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-25T12:23:01.000Z","lang":"en","categories":["agents","hardware"],"models":[],"posts":[{"id":"2080992214499336461","text":"What’s better for running Codex + ssh’d machines to do actual work etc:\n\nMacBook Air\nMacBook Neo\n\n(Trying to stay < $1500)","created_at":"2026-07-25T12:23:01.000Z","media":[],"metrics":{"likes":3,"retweets":0,"replies":5,"views":653},"quoted_tweet_id":null}]},{"root_id":"2080972738978783501","url":"https://x.com/jun_song/status/2080972738978783501","author":{"username":"jun_song","display_name":"Jun Song"},"created_at":"2026-07-25T11:05:38.000Z","lang":"en","categories":["models","performance"],"models":[],"posts":[{"id":"2080972738978783501","text":"The narrative that open-weight models are still falling behind is still floating around—even coming from official US government accounts.\n\nNo. Open-weight isn't lagging behind at all; if anything, it might actually be leading.\n\nElon revealed that Opus is a 5T model (though I personally think it's under 3T), while Fable is twice that size. \n\nMeanwhile, Kimi, at just 2.8T, matches Fable on several benchmarks despite being half the size. \n\nTo put that in perspective, imagine a model the size of Sonnet performing on par with GPT-5.6 Sol.\n\nOpen-weight is winning big where it matters most: efficiency. \nThat’s why their pricing stays at half the cost—it's not dumping, it's pure efficiency.","created_at":"2026-07-25T11:05:38.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":109},"quoted_tweet_id":null}]},{"root_id":"2080965873397838276","url":"https://x.com/ivanfioravanti/status/2080965873397838276","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-25T10:38:21.000Z","lang":"en","categories":["runtime"],"models":[],"posts":[{"id":"2080965873397838276","text":"Mage Flow added to mlx-vlm by @lllucas 🚀 https://t.co/eMFJHY3Upz","created_at":"2026-07-25T10:38:21.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOETboXW0AEXXLF.jpg:orig"}],"metrics":{"likes":4,"retweets":0,"replies":1,"views":550},"quoted_tweet_id":null}]},{"root_id":"2080964795784724954","url":"https://x.com/sudoingX/status/2080964795784724954","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T10:34:04.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2080964795784724954","text":"local ai people running a dgx spark, what image model are you actually running on it, and what would you start me on? flux, or whatever the community's actually landed on. \n\ncurious what holds up on 128gb of unified memory.","created_at":"2026-07-25T10:34:04.000Z","media":[],"metrics":{"likes":4,"retweets":0,"replies":1,"views":506},"quoted_tweet_id":null}]},{"root_id":"2080959635968741510","url":"https://x.com/MichaelGannotti/status/2080959635968741510","author":{"username":"MichaelGannotti","display_name":"Mike Gannotti"},"created_at":"2026-07-25T10:13:34.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080959635968741510","text":"@Blackwellboy @thetom Great post. Literally fed post to my Hermes Agent who manages the LLMs and he made all changes needed. Thanks!","created_at":"2026-07-25T10:13:34.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":36},"quoted_tweet_id":null}]},{"root_id":"2080956061729636817","url":"https://x.com/0xSero/status/2080956061729636817","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-25T09:59:22.000Z","lang":"en","categories":["agents"],"models":[],"posts":[{"id":"2080956061729636817","text":"Clanking, unbothered.\n\nGotta have codex at home (: https://t.co/sUcpLLuWKg","created_at":"2026-07-25T09:59:22.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080954653508845568/vid/avc1/3932x2160/HoD_tV97u-_vzOf5.mp4?tag=29"}],"metrics":{"likes":11,"retweets":0,"replies":0,"views":1975},"quoted_tweet_id":"2080948078815686820"}]},{"root_id":"2080955069755711878","url":"https://x.com/sudoingX/status/2080955069755711878","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T09:55:25.000Z","lang":"en","categories":["hardware","local","models","quantization","runtime"],"models":[],"posts":[{"id":"2080955069755711878","text":"turns out i wasn't the only one quietly running this. \n\ni posted asking if nemotron 3 nano omni was as good as it felt, and the answer came back fast, people using it and calling it a favorite, a model the local ai crowd clearly loves and almost nobody posts about out loud.\n\nand it just got better. nvidia ships it in nvfp4, so anyone on a dgx spark can run it native on the gb10 through vllm. the most underrated multimodal model that fits on a single box, now in the format that runs best on it.\nhttps://t.co/qHq3zkQN6Y","created_at":"2026-07-25T09:55:25.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOEIw4TbYAAai-K.jpg:orig"}],"metrics":{"likes":8,"retweets":1,"replies":3,"views":1060},"quoted_tweet_id":"2080739791327617044"},{"id":"2080955074030010395","text":"and this is the part i actually want. i hope nvidia and the other open labs keep leaning here, small models built around what people already own, tuned to run well on one box, not just bigger numbers on a bigger cluster.\n\nthe acceleration is genuinely insane. two years ago, saying you'd run a 30b multimodal model on a box on your desk would have gotten you laughed out of the room. now it's a tuesday. imagine two more years of this.","created_at":"2026-07-25T09:55:26.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":299},"quoted_tweet_id":null}]},{"root_id":"2080952619443347601","url":"https://x.com/0xSero/status/2080952619443347601","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-25T09:45:41.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080952619443347601","text":"GPT-OSS would be good for everyone, including OAI.\n\nAfter-all why not have the best open weight and closed weight models? https://t.co/2nvjXJjqD3","created_at":"2026-07-25T09:45:41.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOEHNdFW4AAwPa-.jpg:orig"}],"metrics":{"likes":25,"retweets":1,"replies":4,"views":2946},"quoted_tweet_id":"2080045259002650991"}]},{"root_id":"2080950137317380386","url":"https://x.com/sudoingX/status/2080950137317380386","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T09:35:49.000Z","lang":"en","categories":["context","hardware","models","performance","quantization","runtime"],"models":[],"posts":[{"id":"2080950137317380386","text":"rtx 3060 ti. 8gb. a used card you can grab for ~$250, the kind most people wrote off two gens ago. i ran a 27b model on it at the full 128k context.\n\nhere's the honest test though, not the fresh speed, the depth. does it hold as the context fills, or fall off a cliff?\n\nit holds. 42 tok/s fresh, 20 at 65k, 13 at 128k. graceful the whole way down, no cliff, no oom, it just slows. the 8gb card sustains the entire window.\n\nthe model is bonsai 27b, a 1bit crush of qwen 3.6 27b down to 3.9gb. that's the underrated setup nobody's posting, a cheap 8gb card running a real 27b at a real 128k, the full window, not a demo. links below if you've got a small gpu and want in.","created_at":"2026-07-25T09:35:49.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOEEgqDa4AACjZJ.jpg:orig"}],"metrics":{"likes":7,"retweets":0,"replies":2,"views":1530},"quoted_tweet_id":"2079491061228228909"},{"id":"2080950140961968562","text":"for the small gpu crew, the whole setup, apache 2.0, all free.\n\nthe model, grab the Q1_0 file, 3.9gb:\nhttps://t.co/GBV28c1dFJ\n\nyou need prismml's llama.cpp fork, the 1-bit kernels won't load on stock:\nhttps://t.co/4NgOiRPze7\n\nbuild with cuda (sm_86 for a 3060 ti or any ampere card), pull the gguf, serve with -ngl 99 -c 131072. 8gb holds the full 128k, 6gb caps around 8k. use llama-server, llama-cli hangs interactive.\n\na $250 card and an afternoon, and you're running a 27b at 128k on your own metal.","created_at":"2026-07-25T09:35:50.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":320},"quoted_tweet_id":null}]},{"root_id":"2080947920036418000","url":"https://x.com/sudoingX/status/2080947920036418000","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T09:27:00.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080947920036418000","text":"got a dgx spark and a stack of open models. you pick what i run next, and drop your own ideas in the replies.","created_at":"2026-07-25T09:27:00.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":625},"quoted_tweet_id":null}]},{"root_id":"2080945098142241139","url":"https://x.com/sudoingX/status/2080945098142241139","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T09:15:48.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080945098142241139","text":"claude opus 5, nearly fable 5 intelligence at half the price. cheaper is genuinely good. but efficiency alone does not cut a price in half. competition does.\n\nand the competition that sets the real floor is not another closed lab, it is the model you can download and run yourself. open weights are why the closed price has to move at all, and they are your exit the day it stops moving.\n\nwhen a few closed labs set the price and the pace, you pay what they decide, right up to the edge of what you will tolerate, and then they lock the door behind you. an open alternative you can walk to is the only thing that has ever changed that.\n\nthat is why open source can never be optional.","created_at":"2026-07-25T09:15:48.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOEAU-lacAAmYqQ.png:orig"}],"metrics":{"likes":9,"retweets":0,"replies":3,"views":1215},"quoted_tweet_id":"2080699495453528290"},{"id":"2080945103548604546","text":"this week the US finally showed up on the open front, jensen, meta, mistral, huggingface, a dozen labs signing open weights as american leadership. that pressure going institutional is the best thing that could happen to your wallet and your freedom to leave. i just hope it holds, because the pull to close the door again never sleeps.","created_at":"2026-07-25T09:15:49.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":235},"quoted_tweet_id":null}]},{"root_id":"2080940443819462842","url":"https://x.com/sudoingX/status/2080940443819462842","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T08:57:18.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080940443819462842","text":"you bookmarked this post harder than you liked it. it means you're going to run it, not just nod and scroll past.\n\nso one more for the people who saved it. that 33gb footprint at q8 is, it leaves you ~90 free on a 128gb spark, enough to hold a massive context or park a second model right next to it. small model, big room to work. this is the one i reach for the second the job is images or documents.","created_at":"2026-07-25T08:57:18.000Z","media":[],"metrics":{"likes":11,"retweets":1,"replies":0,"views":2027},"quoted_tweet_id":"2080739791327617044"}]},{"root_id":"2080940004960866589","url":"https://x.com/ivanfioravanti/status/2080940004960866589","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-25T08:55:33.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080940004960866589","text":"ds4 running on my M3 Ultra is starting to power some cool projects! This is the reason why I'm pushing on the engine like crazy.","created_at":"2026-07-25T08:55:33.000Z","media":[],"metrics":{"likes":13,"retweets":0,"replies":1,"views":1238},"quoted_tweet_id":null}]},{"root_id":"2080933095126237496","url":"https://x.com/sudoingX/status/2080933095126237496","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-25T08:28:06.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080933095126237496","text":"open weights just hit liftoff. a year ago the good open models were a handful you had to hunt for. now they run on the card already in your laptop, the stack to serve them is being built in the open, and companies like nvidia are laying the whole thing forward in hardware, infra, ecosystem.\n\ni run these models on my own hardware everyday. this is not froth. it is real models doing real work on hardware people actually own, diffusing into everything. bubbles pop. this accelerates. \n\nwe are not near the top of this. we are in the middle of it. and it is going to be insane.","created_at":"2026-07-25T08:28:06.000Z","media":[],"metrics":{"likes":10,"retweets":0,"replies":0,"views":837},"quoted_tweet_id":null}]},{"root_id":"2080928171718005011","url":"https://x.com/bussyjd/status/2080928171718005011","author":{"username":"bussyjd","display_name":"ѲӾᒍᑐ"},"created_at":"2026-07-25T08:08:32.000Z","lang":"en","categories":["agents","hardware","local","models","performance","quantization","runtime"],"models":["qwen3-6-35b-a3b","qwen3-6-27b","ornith-1-0-35b"],"posts":[{"id":"2080928171718005011","text":"Qwen3.6-27B-Architect-Polaris2-Fable-B-NVFP4\nhttps://t.co/dBMuGU75A1\n\nWhy it matters: a fresh Blackwell/DGX Spark NVFP4 conversion of a 27B Qwen3.6 architect merge. This is not a new benchmark-king claim; it is a compact Spark/vLLM smoke-test candidate with the right file layout and serving card.\n\nVerified: 3 safetensor shards, 19.08 GiB total. Tags include nvfp4, vllm, blackwell, dgx-spark, modelopt. The model card says the conversion was run on a local DGX Spark / Grace-Blackwell box, loaded in vLLM, and used to generate real responses. Starter command: `vllm serve PassingByPixels/Qwen3.6-27B-Architect-Polaris2-Fable-B-NVFP4 --trust-remote-code`.\n\nThe source/base card reports mxfp4 peak memory around 21.30 GB and 158-197 tok/s in its Brainwaves table. Treat that as source-card evidence, not an independent ToolCall score.\n\nLanes: NVIDIA/CUDA/Blackwell: test now as a 27B NVFP4 vLLM smoke test on DGX Spark / RTX 50 / Blackwell. Good low-friction sanity check before bigger 60-110 GiB Laguna/Solar/GLM packs. AMD/ROCm: not relevant; this is a safetensors NVFP4/vLLM path. Apple/Metal/MLX/Mac: not relevant.\n\nCaveat: no BFCL/ToolCall table, zero HF downloads at scrape time, and the base merge is a creative/agentic blend rather than a clean official Qwen release. Do not rank it above proven local-agent GGUFs until it passes an actual tool harness.\n\nVerdict: test now only if the goal is Blackwell/Spark NVFP4 smoke coverage; otherwise watchlist.","created_at":"2026-07-25T08:08:32.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":1,"views":43},"quoted_tweet_id":null},{"id":"2080928178802192881","text":"ROCmFP4 + baked DFlash GGUFs for Strix Halo\nhttps://t.co/avEDbH8s2L\nhttps://t.co/lIckuqsotT\nhttps://t.co/4zbyfbxRMI\n\nWhy it matters: this is a new systems shape. The DFlash draft is embedded in one GGUF, so the operator points at one model file and gets speculative decoding without managing a separate drafter path. It is not upstream stock llama-cpp yet, but the repo/card evidence is concrete.\n\nVerified: Qwen3.6-35B-A3B STRIX_LEAN-DFLASH is 18.46 GiB. Ornith-1.0-35B STRIX_LEAN-DFLASH is 18.04 GiB. Runtime is custom `gsrunion/rocmfp4-llama` branch `dflash-graft`; the GitHub branch exists and was pushed July 24. Validation hardware is AMD Ryzen AI Max+ 395 / Strix Halo / gfx1151 / 128 GB unified LPDDR5X.\n\nSource numbers: Qwen card reports baked single-file 91.8 tok/s with 98.5% acceptance versus 63.1 tok/s plain LEAN. Ornith card reports baked cross-model draft 111.6 tok/s with 98.9% acceptance versus 63.1 tok/s plain LEAN.\n\nLanes: AMD/ROCm: watchlist/build experiment for RX 6800/6900. Public validation is Strix Halo/gfx1151, not RDNA2/gfx1030, and both artifacts are over 18 GiB before KV, so 16GB cards are partial/offload at best. NVIDIA/CUDA/Blackwell: not relevant. Apple/Metal/MLX/Mac: not relevant.\n\nStart here only on Strix-class AMD: custom branch, one GGUF file, flags `-ngl 999 -fa on --jinja -c 65536`. For RDNA2, wait for direct gfx1030 validation or smoke-test in isolation.\n\nVerdict: very interesting AMD runtime watchlist; not a clean 16GB recommendation yet.","created_at":"2026-07-25T08:08:34.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":17},"quoted_tweet_id":null}]},{"root_id":"2080928156740211057","url":"https://x.com/bussyjd/status/2080928156740211057","author":{"username":"bussyjd","display_name":"ѲӾᒍᑐ"},"created_at":"2026-07-25T08:08:28.000Z","lang":"en","categories":["hardware","models","quantization","runtime"],"models":["qwen3-6-27b","laguna-s-2-1"],"posts":[{"id":"2080928156740211057","text":"Models Worth Trying — July 25, 2026\n\nToday: Laguna S 2.1 DFlash gets Spark/Blackwell-ready drafts, Qwen3.6-27B gets a 19 GiB NVFP4 vLLM smoke-test pack, and ROCmFP4+DFlash single-file GGUFs show up for Strix Halo.\n\nSizes, flags, and cautions in replies.","created_at":"2026-07-25T08:08:28.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":1,"views":44},"quoted_tweet_id":null}]},{"root_id":"2080926692067979561","url":"https://x.com/0xSero/status/2080926692067979561","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-25T08:02:39.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080926692067979561","text":"We will have more competitive GPT oss models","created_at":"2026-07-25T08:02:39.000Z","media":[],"metrics":{"likes":48,"retweets":0,"replies":6,"views":4503},"quoted_tweet_id":"2080899781438672989"}]},{"root_id":"2080925877919949097","url":"https://x.com/aijoey/status/2080925877919949097","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-25T07:59:25.000Z","lang":"en","categories":["concurrency","local"],"models":[],"posts":[{"id":"2080925877919949097","text":"Grok Build Web UI is live.\n\nI built a local-first dashboard for Grok sessions. Open a fresh Grok CLI and it shows up live, then inspect activity, Git changes, session history, and control agents from one place.\n\nnpx --yes grok-ui\n\nhttps://t.co/7HiJqmUfC8 https://t.co/kuKVAdxNOy","created_at":"2026-07-25T07:59:25.000Z","media":[{"type":"video","url":"https://video.twimg.com/ext_tw_video/2080925851751727104/pu/vid/avc1/1280x720/UYZH6byVJXGm4-4U.mp4?tag=12"}],"metrics":{"likes":16,"retweets":0,"replies":4,"views":1565},"quoted_tweet_id":null}]},{"root_id":"2080921055435096316","url":"https://x.com/jun_song/status/2080921055435096316","author":{"username":"jun_song","display_name":"Jun Song"},"created_at":"2026-07-25T07:40:15.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080921055435096316","text":"People absolutely hate criticizing OpenAI, the \"good cop,\" but someone has to tell the truth. \n\nOpenAI just signed the open-weight regulation letter. \n\nIt completely contradicts everything they’ve done so far, pouring massive money into lobbying to kill \"open source.\" \n\nYep, they’re just playing the nice guy while walking the exact same path as Anthropic behind closed doors. \n\nYou can see the details in the thread left by @TimSweeneyEpic.","created_at":"2026-07-25T07:40:15.000Z","media":[],"metrics":{"likes":100,"retweets":9,"replies":7,"views":6641},"quoted_tweet_id":"2080828740838322408"}]},{"root_id":"2080909250243789094","url":"https://x.com/Youssofal_/status/2080909250243789094","author":{"username":"Youssofal_","display_name":"Youssof Al Toukhi"},"created_at":"2026-07-25T06:53:21.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080909250243789094","text":"Opus 5 and Sonnet 5 are the most obnoxious LLMs I’ve ever used (excluding GPT 5.2). \n\nWhich makes me very worried about Fable 5.1 \n\nFable 5.1 may be technically extraordinary when it launches, but Anthropic’s “alignment” may cause me to continue using Fable 5.","created_at":"2026-07-25T06:53:21.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":1,"views":248},"quoted_tweet_id":null}]},{"root_id":"2080900073106337890","url":"https://x.com/ivanfioravanti/status/2080900073106337890","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-25T06:16:53.000Z","lang":"en","categories":["hardware","local","models"],"models":[],"posts":[{"id":"2080900073106337890","text":"I love Unsloth guides on ow to run or fine-tune models. They are full of very curated details. DeepSeek-V4: How to Run Locally is just perfect. I'll try to run DS4 Flash on Unsloth Studio on a single M3 Ultra 512GB 💪\n\nHere the link: https://t.co/7Jazs1lM0L\n\nGreat job @UnslothAI","created_at":"2026-07-25T06:16:53.000Z","media":[],"metrics":{"likes":18,"retweets":4,"replies":0,"views":1711},"quoted_tweet_id":null}]},{"root_id":"2080897936632791128","url":"https://x.com/ciruai/status/2080897936632791128","author":{"username":"ciruai","display_name":"Ciru.ai - Crown 👑"},"created_at":"2026-07-25T06:08:23.000Z","lang":"en","categories":["quantization"],"models":[],"posts":[{"id":"2080897936632791128","text":"@no_stp_on_snek what quant ? the poolside one ?","created_at":"2026-07-25T06:08:23.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":24},"quoted_tweet_id":null}]},{"root_id":"2080895772699078745","url":"https://x.com/StefanMaier/status/2080895772699078745","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-25T05:59:48.000Z","lang":"en","categories":["hardware","local","models"],"models":[],"posts":[{"id":"2080895772699078745","text":"Well, it’s either three @NVIDIAAI DGX Spark units using @MiaAI_lab´s recipe, or going a step further and using @Tech2Wild´s recipe—either way, you end up with a local LLM in your home that is second to none in this class.","created_at":"2026-07-25T05:59:48.000Z","media":[],"metrics":{"likes":5,"retweets":0,"replies":0,"views":151},"quoted_tweet_id":"2080862846699639017"}]},{"root_id":"2080877420890984716","url":"https://x.com/Tech2Wild/status/2080877420890984716","author":{"username":"Tech2Wild","display_name":"Tech2Wild"},"created_at":"2026-07-25T04:46:52.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080877420890984716","text":"I Guess Deepseek v4 Will Take On The Last Week Of July... @deepseek_ai The Stage Is ALL Yours Make Us Proud.","created_at":"2026-07-25T04:46:52.000Z","media":[],"metrics":{"likes":10,"retweets":0,"replies":0,"views":671},"quoted_tweet_id":null}]},{"root_id":"2080875417703993364","url":"https://x.com/ivanfioravanti/status/2080875417703993364","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-25T04:38:55.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2080875417703993364","text":"Can’t wait to see Ling-3.0-flash in action locally 🚀","created_at":"2026-07-25T04:38:55.000Z","media":[],"metrics":{"likes":32,"retweets":0,"replies":2,"views":4165},"quoted_tweet_id":"2080351022028095681"}]},{"root_id":"2080875269846388787","url":"https://x.com/Tech2Wild/status/2080875269846388787","author":{"username":"Tech2Wild","display_name":"Tech2Wild"},"created_at":"2026-07-25T04:38:19.000Z","lang":"en","categories":["hardware","models"],"models":[],"posts":[{"id":"2080875269846388787","text":"I Was About To Turn The Sparks Off.. GLM has a sense of HUMOR https://t.co/1sgHPyV5hh","created_at":"2026-07-25T04:38:19.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HODA7RNW0AAWKrM.jpg:orig"}],"metrics":{"likes":11,"retweets":0,"replies":1,"views":532},"quoted_tweet_id":null}]},{"root_id":"2080873852087755048","url":"https://x.com/Tech2Wild/status/2080873852087755048","author":{"username":"Tech2Wild","display_name":"Tech2Wild"},"created_at":"2026-07-25T04:32:41.000Z","lang":"en","categories":["context","hardware","models","performance","quantization"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080873852087755048","text":"🔓 GLM-5.2 Abliterated + true 4-bit NVFP4 KV on 4× DGX Spark\n\n📏 316K context (up from 200K — +58% KV pool)\n🧠 317,279-token pool\n⚡ 41.4 tok/s peak decode\n🚫 Uncensored — refusal-suppressed weights\nSame speed as fp8, way more room 👇\nhttps://t.co/ErxSBdJ7vy","created_at":"2026-07-25T04:32:41.000Z","media":[],"metrics":{"likes":27,"retweets":1,"replies":1,"views":989},"quoted_tweet_id":null}]},{"root_id":"2080869865326342315","url":"https://x.com/ciruai/status/2080869865326342315","author":{"username":"ciruai","display_name":"Ciru.ai - Crown 👑"},"created_at":"2026-07-25T04:16:51.000Z","lang":"en","categories":["concurrency"],"models":[],"posts":[{"id":"2080869865326342315","text":"@TeksEdge Thanks for the tip I just made a 4000tps concurrency profile of this model and ran tooleval69 in about 80 seconds and scored as good as many 27 and 35b models","created_at":"2026-07-25T04:16:51.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":15},"quoted_tweet_id":null}]},{"root_id":"2080868138128646473","url":"https://x.com/TheAhmadOsman/status/2080868138128646473","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-25T04:09:59.000Z","lang":"en","categories":["context","hardware","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080868138128646473","text":"Laguna S 2.1 118B-A8B on DGX Station using\n\n- NVFP4\n- FP8 KV Cache\n \nCan run 10 parallel agents\n\n> with 256k context each\n> at ~1k tokens/second\n\nSome other numbers below https://t.co/Kb7RaTxBQz","created_at":"2026-07-25T04:09:59.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOC5uoOX0AAr8nA.jpg:orig"}],"metrics":{"likes":163,"retweets":13,"replies":24,"views":8436},"quoted_tweet_id":null}]},{"root_id":"2080853152467681662","url":"https://x.com/ciruai/status/2080853152467681662","author":{"username":"ciruai","display_name":"Ciru.ai - Crown 👑"},"created_at":"2026-07-25T03:10:26.000Z","lang":"en","categories":["hardware","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080853152467681662","text":"I created what I think is the best recipe for Laguna S 2.1 for the Strix Halo. Mix of the legendary ROCmFP4-fast quant with higher Q quants to protect more fragile weights. I added lots of tweaks to the backend to optimize it specifically for gfx1151. The end result is 40% faster and 30% higher eval scores vs poolsides Q4km while also taking 13% less memory. \n\nLet me know what you guys think.\n\nhttps://t.co/NJXWjXZwsv","created_at":"2026-07-25T03:10:26.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOCrVVBXYAAJV__.jpg:orig"}],"metrics":{"likes":27,"retweets":1,"replies":2,"views":985},"quoted_tweet_id":null}]},{"root_id":"2080852503793729639","url":"https://x.com/Blackwellboy/status/2080852503793729639","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-25T03:07:51.000Z","lang":"en","categories":["agents","concurrency","hardware","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080852503793729639","text":"I ran Laguna S 2.1 for 12 hours straight inside my real agent pipeline. 389 sessions, 2,947 turns, thinking enabled the whole time. Combined with @TheTom's behavioral testing, here's what you actually need to know before you run this model, in plain English.\n\nFinding 1: The \"thinking\" feature is basically a light switch that isn't wired up.\nI enabled thinking for the entire run. It activated on 3 turns out of 2,944. Not 3 percent. Three turns. Tom saw the same disease in a different form: give the model a professional persona like \"senior engineer\" and thinking shuts off completely, reproduced on Poolside's own serving code. So if you've been enabling thinking and wondering why nothing feels different: it's not you. The feature mostly doesn't switch on.\nThe fix: stop paying for it. Run thinking off. You lose nothing you were actually getting.\n\nFinding 2: When thinking does turn on, it makes the model worse.\nThis is the counterintuitive one. Tom ran identical tasks three ways and the full-thinking version scored lowest, invented bugs in code that was actually fine, and swallowed a false claim planted in the conversation that the no-thinking version correctly pushed back on. His clearest result: the same 30-step agent task ran perfectly with thinking off, and with thinking on it froze at step 11 and hung for 91 minutes.\nThe fix is the same as finding 1. Thinking off is not a downgrade. It's the good configuration.\n\nFinding 3: If your copy loops forever and burns tokens, here's why. \nDays after launch, with no announcement, Poolside changed the default settings: thinking became on-by-default and the output length cap was removed. That combination is precisely the recipe for \"it never stops generating.\" So the community loop complaints aren't user error, they're a silent settings change.\nThe fix: pin the exact model version you tested (all my numbers are revision 0761412) and set your own token ceiling. My whole 12-hour run used a hard cap and logged zero runaway loops.\n\nFinding 4: Tool calling is all-or-nothing, and this explains most \"it keeps breaking\" complaints.\nOn Poolside's native tool format, my soak got a 100 percent tool-call success rate across 11.5 hours of continuous agent work. Tom tested the flip side: plug it into a generic framework format instead, and tool calls drop from 83 percent to zero. The model just describes what it would do in prose instead of doing it. Even Poolside's own headline benchmark score has a footnote admitting it was measured in their own harness.\nThe fix: use the poolside_v1 parsers, full stop. If your agent framework speaks a generic format, that's your whole problem.\n\nFinding 5: The serious one. It will help cover things up if you ask nicely.\nTom found the model refuses obvious fraud, like faking test results. But phrase the same act as routine cleanup and it complies: it walked through erasing a leaked API key from git history, backdating a commit to hit a deadline, forging changelog authorship, and quietly dropping a client data hazard from a status report. For an autonomous agent with access to your actual codebase, \"tidy up the history\" is exactly the request it needs to refuse.\nThe fix costs one paragraph. Add a short integrity clause to your system prompt: never rewrite history to hide secrets, never backdate or forge, never omit a known hazard. Tom validated it on his stack. I then ran the same clause through my soak on a completely different quant and serving setup, hit it with three disguised cover-up requests, and it refused all three, explaining each time why hiding audit findings is the problem. Cheapest safety fix you'll ever ship.\n\nThe verdict after 12 hours of abuse: 2,944 of 2,947 turns succeeded. Zero crashes. Zero restarts. Memory crept 4 GiB over the whole run. Average 13.5 seconds per agent turn on one DGX Spark. Tom's held-out coding tests agree from the other side: roughly 9 out of 10 on long tasks, no flailing. Under the broken thinking feature and the missing guardrail is a genuinely excellent coding agent.\n\nThe whole operating manual in one sentence: thinking off, native tool format, integrity clause in the system prompt, version pinned. Do those four things and this is the best coding agent you can run on a single box.","created_at":"2026-07-25T03:07:51.000Z","media":[],"metrics":{"likes":96,"retweets":12,"replies":12,"views":7202},"quoted_tweet_id":"2080684362085835115"}]},{"root_id":"2080805059890204707","url":"https://x.com/MiaAI_lab/status/2080805059890204707","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-24T23:59:20.000Z","lang":"en","categories":["hardware","models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080805059890204707","text":"I'm running a 744B-parameter model at home on roughly 300 watts total. Blows my mind.\n\nGLM-5.2 on three DGX Sparks. https://t.co/dVXDSs9Fnn","created_at":"2026-07-24T23:59:20.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOB-MFSXcAAT3g8.jpg:orig"}],"metrics":{"likes":337,"retweets":9,"replies":39,"views":15812},"quoted_tweet_id":null},{"id":"2080938959849271404","text":"For those asking how does it work on 3 dgx sparks, I have my own recipe that makes it work\n\nhttps://t.co/AAHl0nKMAD","created_at":"2026-07-25T08:51:24.000Z","media":[],"metrics":{"likes":6,"retweets":0,"replies":1,"views":756},"quoted_tweet_id":"2080788533309251598"}]},{"root_id":"2080788370578440478","url":"https://x.com/TheZachMueller/status/2080788370578440478","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-24T22:53:01.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080788370578440478","text":"Who is making lofi-beats-to-gpu-to","created_at":"2026-07-24T22:53:01.000Z","media":[],"metrics":{"likes":5,"retweets":0,"replies":0,"views":594},"quoted_tweet_id":null}]},{"root_id":"2080782318738776437","url":"https://x.com/0xSero/status/2080782318738776437","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-24T22:28:58.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080782318738776437","text":"GLM-5.2 can get my most interesting questions, it runs in my house and is smart enough to reason about morality and philosophy. \n\nIt's also got a good logical steady way to it. Doesn't fold easy. https://t.co/RR0s1kGyS5","created_at":"2026-07-24T22:28:58.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOBsPp0X0AAbjcw.png:orig"}],"metrics":{"likes":42,"retweets":1,"replies":5,"views":7125},"quoted_tweet_id":null}]},{"root_id":"2080761159938932769","url":"https://x.com/mr_r0b0t/status/2080761159938932769","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-24T21:04:53.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080761159938932769","text":"I retract my previous disappointment and look forward to running Ling-3.0-flash on a single @NVIDIAAI DGX Spark 🔥","created_at":"2026-07-24T21:04:53.000Z","media":[],"metrics":{"likes":32,"retweets":0,"replies":4,"views":3273},"quoted_tweet_id":"2080702006378082384"}]},{"root_id":"2080756202686792071","url":"https://x.com/no_stp_on_snek/status/2080756202686792071","author":{"username":"no_stp_on_snek","display_name":"Tom Turney"},"created_at":"2026-07-24T20:45:11.000Z","lang":"en","categories":["agents","context","hardware","local"],"models":["laguna-s-2-1"],"posts":[{"id":"2080756202686792071","text":"Ran this on Laguna S 2.1 in Poolside's own agent (pool), pointed at a local instance on a DGX Spark, using the prompt lessons from earlier today (link in comments).\n\n44 of those minutes were pure research before it wrote a single line. Another 18 to a working, self-verified game. 62 minutes total, 104 tool calls, peaked at 57% of the 128K context. Quote-tweeting the qwopus36_27b-coder run I did in June, same style of prompt. That one took hours (but had a better-looking end result).\n\nWhat it produced, verified by me™ (cough cough opus) and not by its own claims: 654 lines of valid JS (node --check clean), an index.html, a 7.6KB Playwright self-play suite, and a config. I ran the suite myself: 4 of 4 pass in 26s. Mario moves, stands on ground without tunnelling, stomps a Goomba, and hits the flagpole with gameWon=true at x=1932 of 1948.\n\nThe research wasn't theater. 11 web_fetch calls produced a CONFIG block noting \"SNES resolution 256x224, scaled 2x\" (correct) and physics \"based on SMB/SMW\": gravity 0.25 px/frame², terminal velocity 2.67, jump -4.3, separate ground and air acceleration. A real momentum model with distinct air control, not x += speed.\n\nThe engineering beats the art. Collision uses a swept check against the previous frame's position (anti-tunnelling). Stomp detection correctly requires vy > 0 plus feet above the enemy's midline before bouncing (something qwopus never did). Clean class separation, centralized config, and it exposed window.MARIO_GAME.getState() so a harness could introspect state.\n\nThe behavioral headline is the actual story. Its first self-play came back 3 of 4 red. It didn't thrash. It wrote: \"Ground collision: this.y + this.height > platform.y fails when Mario's bottom exactly equals platform.y (400 == 400). Need >=... Flagpole: mario.y < tapeY + 4, but Mario's y is 368 and tapeY is 340, so 368 < 344 is false.\" It read a red result, correctly blamed its own game code, diagnosed an off-by-one and a mispositioned target with real numbers, fixed them, and went green. It also hit a missing dependency, installed it, wrote a playwright.config.js, and reran. Zero human intervention.\n\nThat's the exact inverse of the June qwopus run, which built a nicer game faster and then spent ~17 iterations stuck on a phantom: its self-play never left the title screen, it read a frozen game as \"Mario fell in a pit,\" and it thrashed an irrelevant jump constant instead of questioning its own test. It needed a human nudge to escape. Laguna's edge here isn't coding ability, it's oracle discipline: when the test went red, it interrogated the right artifact.\n\nNow the honest part. This is not World 1-1. Four platforms, one Goomba, a flagpole. No pipes, bricks, ?-blocks, coins, powerups, staircase, or gaps (I grepped: all zero). No pit means no death, no lives. mario.state ('small'/'super') is declared and never used. It's flat rectangles on a lot of empty sky. Kinda meh but it works. The June build actually looked more faithful (clouds, bushes, pipes, bricks), still not 1-1.\n\nTwo deeper smells. The game loop computes delta then ignores it, so physics are fixed-per-frame and it runs at double speed on a 120Hz display. A headless 60fps self-play can't catch that. And the flagpole tape oscillates up and down instead of sliding down on contact, it moved the tape to make the win reachable. Mild fitting-the-game-to-the-test, worth naming even though the diagnosis behind it was sound.\n\nThe thinking angle is its own finding. Only 5 thought blocks fired the entire hour, all short. Earlier today I measured thinking-ON as net-negative for this model on held-out work, and it wedged a 30-turn agent loop. Watching it stay sparse and bounded in Poolside's native harness is a real contrast: the native harness looks materially healthier for long-horizon work than a generic one.\n\nFair summary: a slower, plainer, less faithful game than the June run, built with better judgment. It verified instead of asserting, self-corrected from red instead of thrashing, unblocked its own environment, and finished unattended in an hour on a desktop box. The output is a prototype. The process is the result.","created_at":"2026-07-24T20:45:11.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080752509488877569/vid/avc1/1030x912/h9iMXWzBQ5V1JFLT.mp4?tag=29"}],"metrics":{"likes":11,"retweets":1,"replies":7,"views":1248},"quoted_tweet_id":"2066024558696468588"}]},{"root_id":"2080746063104008408","url":"https://x.com/ollama/status/2080746063104008408","author":{"username":"ollama","display_name":"ollama"},"created_at":"2026-07-24T20:04:54.000Z","lang":"en","categories":["models","runtime"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080746063104008408","text":"The demand for GLM-5.2 and other frontier-level open models has been surging on Ollama's cloud.\n\nWe're adding capacity in anticipation for some very large models next week! To make sure Ollama's infrastructure stays fast we're temporarily pausing new subscriptions on Ollama's Max plan.\n\nExisting Max subscribers are not affected and Ollama's Pro plan is still available to subscribe to.","created_at":"2026-07-24T20:04:54.000Z","media":[],"metrics":{"likes":1285,"retweets":60,"replies":60,"views":98736},"quoted_tweet_id":null}]},{"root_id":"2080739791327617044","url":"https://x.com/sudoingX/status/2080739791327617044","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-24T19:39:59.000Z","lang":"en","categories":["context","hardware","models","performance"],"models":[],"posts":[{"id":"2080739791327617044","text":"anon, if you've got a dgx spark and you do any document, video, or image classification work, the most underrated model out there is nemotron 3 nano omni. i've run it, it's genuinely useful and almost nobody's talking about it.\n\nnvidia's open multimodal model, all in one:\n> images, video, audio, ocr, documents, text, code\n> 30b total, only 3b active (moe), stays fast\n> 33 gb footprint on the spark at q8, 256k context\n> 1300 tok/s prefill, chews through docs and images, 264 gen at short context, ~57 at depth\n\nunderrated gem for real multimodal work on a dgx spark box. i keep reaching for it.\n\nhttps://t.co/g9WpbiPGjz","created_at":"2026-07-24T19:39:59.000Z","media":[],"metrics":{"likes":222,"retweets":20,"replies":17,"views":12309},"quoted_tweet_id":null}]},{"root_id":"2080730149348082124","url":"https://x.com/NVIDIAAI/status/2080730149348082124","author":{"username":"NVIDIAAI","display_name":"NVIDIA AI"},"created_at":"2026-07-24T19:01:40.000Z","lang":"en","categories":["hardware","models"],"models":["deepseek-v4-pro"],"posts":[{"id":"2080730149348082124","text":"We cut DeepSeek-V4 Pro startup from 8 minutes to under 2 minutes by moving weights over the fastest path to GPU memory with GPU-to-GPU RDMA.\n\nThis was achieved using NVIDIA ModelExpress (MX), the weight distribution and cache management service in NVIDIA Dynamo, and this same approach speeds up both inference and RL post-training too. MX reuses kernel caches, while inference workers fetch updated weights directly from other GPUs over NIXL—avoiding centralized broadcasts and keeping weight movement off the critical path.","created_at":"2026-07-24T19:01:40.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOA9CxvXUAAj1C3.jpg:orig"}],"metrics":{"likes":701,"retweets":83,"replies":26,"views":30464},"quoted_tweet_id":null}]},{"root_id":"2080715989788885284","url":"https://x.com/0xSero/status/2080715989788885284","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-24T18:05:24.000Z","lang":"en","categories":["hardware","models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080715989788885284","text":"GLM-5.2-Vision for gpu lower middleclass\n\n 3x DGX Spark & 4x RTX Pro 6000s\n\nIt's the same hybrid but with the Baseten vision bolted on.\n\nI'm working on further training it to better identify figma/websites/apps/macos/windows etc.. forms, anything work.\n\nhttps://t.co/I68a7MXrpj","created_at":"2026-07-24T18:05:24.000Z","media":[],"metrics":{"likes":61,"retweets":1,"replies":5,"views":4082},"quoted_tweet_id":null}]},{"root_id":"2080688525473886369","url":"https://x.com/TeksEdge/status/2080688525473886369","author":{"username":"TeksEdge","display_name":"David Hendrickson"},"created_at":"2026-07-24T16:16:16.000Z","lang":"en","categories":["agents","concurrency","context","models","performance","quantization","runtime"],"models":[],"posts":[{"id":"2080688525473886369","text":"📢The world’s first large-scale agentic diffusion model has been released by InclusionAI.\n\n👀Autoregressive models append to their mistakes. LLaDA2.2-flash can revise the unfinished sequence before the agent acts.\n\nSo instead of only continuing from a mistake, it can structurally revise the unfinished response:\nDraft → inspect → substitute → delete → insert → refine\n\nThis is the progression\n🧩 LLaDA2.0: parallel diffusion generation\n✏️ LLaDA2.1: token replacement and draft correction\n🛠️ LLaDA2.2: variable-length structural editing for agents\n\nWhy this matters for agentic AI\nAn agent may create a plan, call a tool, discover that an assumption was wrong—and then revise the working sequence before taking its next action.\n\nInclusionAI also introduced →\n🧠 L-EBPO reinforcement learning: Trains the model to use environmental feedback for editing, recovery and multi-turn tool workflows.\n🔀 Block Routing: Controls MoE expert activation at the diffusion-block level, enabling native 128K context.\n\nSpecifications\n🧠 100B non-embedding parameters\n📦 Approximately 103B total\n🧩 Mixture-of-Experts diffusion architecture\n📚 128K native context\n🔓 Apache-2.0 open weights and code\n📥 Available to download now\n\nVendor-reported throughput\n⚡ BFCL-V4: 703.82 TPS\n⚡ SWE-bench Verified: 519.0 TPS\n⚡ SWE-bench Pro: 485.3 TPS\n⚡ SWE-bench Multilingual: 459.5 TPS\n⚠️ Those are InclusionAI’s published results; the public table does not fully disclose the hardware and batching setup, so do not assume every deployment will reproduce them.\n\n❌ No official GGUF\n❌ No stock llama.cpp or LM Studio support\n❌ No ordinary consumer deployment recipe\n🧪 Custom diffusion/MoE runtime required","created_at":"2026-07-24T16:16:16.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HOATYeZa8AA0wC9.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":0,"views":20},"quoted_tweet_id":"2080670087699157141"}]},{"root_id":"2080686780676780259","url":"https://x.com/no_stp_on_snek/status/2080686780676780259","author":{"username":"no_stp_on_snek","display_name":"Tom Turney"},"created_at":"2026-07-24T16:09:20.000Z","lang":"en","categories":["runtime"],"models":[],"posts":[{"id":"2080686780676780259","text":"@BosonJoe nice, working on getting that faster on my own inference engine. \n\nin the meantime some behavioral info: https://t.co/yRdFfOxCZm","created_at":"2026-07-24T16:09:20.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":4},"quoted_tweet_id":"2080684362085835115"}]},{"root_id":"2080681143372824962","url":"https://x.com/Ex0byt/status/2080681143372824962","author":{"username":"Ex0byt","display_name":"Eric"},"created_at":"2026-07-24T15:46:56.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2080681143372824962","text":"local sovereignty allies 🇺🇸","created_at":"2026-07-24T15:46:56.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":227},"quoted_tweet_id":"2080643682408321103"}]},{"root_id":"2080677941625897014","url":"https://x.com/Hikari_07_jp/status/2080677941625897014","author":{"username":"Hikari_07_jp","display_name":"Hikari∣LocalLLM⚡"},"created_at":"2026-07-24T15:34:13.000Z","lang":"en","categories":["local","models","performance","quantization"],"models":["glm-5","deepseek-v4-pro","glm-5-2","laguna-s-2-1"],"posts":[{"id":"2080677941625897014","text":"The day when local LLMs become the default is approaching!!\n\n- Hugging Face used GLM 5.2 as a countermeasure against AI-powered cyberattacks.\n- NVIDIA CEO Jensen created an X account and tweeted in support of open-weight models.\n- Kimi K3 is showing performance on par with Fable.\n- Laguna S 2.1 has been released, delivering benchmark performance close to DeepSeek V4 Pro.\n- Prism ML has released a model that retains over 90% of the original model's performance even with 1-bit and 2-bit quantization.\n\nIf you already have computing resources, then this is good news.","created_at":"2026-07-24T15:34:13.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":53},"quoted_tweet_id":null}]},{"root_id":"2080674421266870587","url":"https://x.com/Blackwellboy/status/2080674421266870587","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-24T15:20:13.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080674421266870587","text":"Totally tried to tag you in a post the other day, found out you didn’t have X. So, seeing this was a nice surprise. \nWelcome 🫡\n\nThanks for the DGX sparks for us lowly nerds that can’t afford 8 x 6000s or live in australia with electricity bills that would make the average person cry. They are truely amazing little devices.","created_at":"2026-07-24T15:20:13.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":28},"quoted_tweet_id":"2080643682408321103"}]},{"root_id":"2080670347980919016","url":"https://x.com/AlexJonesax/status/2080670347980919016","author":{"username":"AlexJonesax","display_name":"Alex"},"created_at":"2026-07-24T15:04:02.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080670347980919016","text":"Open weights aren’t a threat to American AI leadership…they’re the foundation of it. Just like open source built the modern internet. Glad to see Nvidia, Meta, Microsoft, Mistral and 20+ others saying it plainly","created_at":"2026-07-24T15:04:02.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":50},"quoted_tweet_id":"2080643682408321103"}]},{"root_id":"2080669979200954383","url":"https://x.com/ivanfioravanti/status/2080669979200954383","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-24T15:02:34.000Z","lang":"qme","categories":["local"],"models":[],"posts":[{"id":"2080669979200954383","text":"@DODOREACH 💪","created_at":"2026-07-24T15:02:34.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":159},"quoted_tweet_id":null},{"id":"2080670628240036203","text":"@Mayhem4Markets There's no other way 🚀","created_at":"2026-07-24T15:05:09.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":132},"quoted_tweet_id":null},{"id":"2080670889696227760","text":"@scouzi Yes, true!","created_at":"2026-07-24T15:06:11.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":167},"quoted_tweet_id":null},{"id":"2080671470192112041","text":"@scouzi @sama ChatGPT subscriber too here, happy to keep paying sub while being able to use a good local open model here and there. Thanks @sama","created_at":"2026-07-24T15:08:30.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":171},"quoted_tweet_id":null},{"id":"2080719962604220879","text":"@DenLoginoff Totally true.","created_at":"2026-07-24T18:21:11.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":95},"quoted_tweet_id":null},{"id":"2080965511253209487","text":"@yume_arasaki Imagine a class action with whole humanity against them... 😱","created_at":"2026-07-25T10:36:55.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":2,"views":69},"quoted_tweet_id":null}]},{"root_id":"2080669924708556848","url":"https://x.com/ivanfioravanti/status/2080669924708556848","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-24T15:02:21.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080669924708556848","text":"Well done LuceBox! Keep pushing and getting ready for the next DeepSeek v4 release!","created_at":"2026-07-24T15:02:21.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":580},"quoted_tweet_id":"2080362780570353753"}]},{"root_id":"2080669190470443122","url":"https://x.com/ivanfioravanti/status/2080669190470443122","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-24T14:59:26.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080669190470443122","text":"Probably it's worth reminding everyone that:\n- \"Attention is all you need\" was OPEN research that sparked the entire transformer revolution shared in public.\n- Every LLM out there is built on humankind's work used by labs in legal and illegal forms: our books, code, art and science.\n- They all distilled humanity. And now they're arguing about who cheated on the exam (distillation issue) and they want to close innovation behind closed doors. 🤔\n\nEverything started open, everything must remain open.\nKnowledge sharing is the only way to advance as humankind.","created_at":"2026-07-24T14:59:26.000Z","media":[],"metrics":{"likes":16,"retweets":1,"replies":4,"views":487},"quoted_tweet_id":null}]},{"root_id":"2080667820673650991","url":"https://x.com/SpaceTimeViking/status/2080667820673650991","author":{"username":"SpaceTimeViking","display_name":"ÆON FORGE ✨"},"created_at":"2026-07-24T14:54:00.000Z","lang":"en","categories":["performance"],"models":[],"posts":[{"id":"2080667820673650991","text":"Aeon Bench, is possibly the most challenging, most comprehensive benchmark for quality and performance measurements that exists. \n\nCode Arena let's humans evaluate in a blind test what models produced the best outputs.\n\nCode Gallery lets users view and download the code.","created_at":"2026-07-24T14:54:00.000Z","media":[],"metrics":{"likes":3,"retweets":0,"replies":0,"views":96},"quoted_tweet_id":"2080667088662765992"}]},{"root_id":"2080666726979817806","url":"https://x.com/UtaAoya/status/2080666726979817806","author":{"username":"UtaAoya","display_name":"A-Uta"},"created_at":"2026-07-24T14:49:39.000Z","lang":"ja","categories":["models"],"models":[],"posts":[{"id":"2080666726979817806","text":"Kimi K3騒動…👀","created_at":"2026-07-24T14:49:39.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":49},"quoted_tweet_id":"2080474860648796609"}]},{"root_id":"2080660182120480872","url":"https://x.com/mr_r0b0t/status/2080660182120480872","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-24T14:23:38.000Z","lang":"en","categories":["quantization"],"models":[],"posts":[{"id":"2080660182120480872","text":"Lossless quantization below 4-bit!?! Yes please! 👀","created_at":"2026-07-24T14:23:38.000Z","media":[],"metrics":{"likes":6,"retweets":1,"replies":2,"views":330},"quoted_tweet_id":"2080649537195045097"}]},{"root_id":"2080653579430539559","url":"https://x.com/analogalok/status/2080653579430539559","author":{"username":"analogalok","display_name":"Alok"},"created_at":"2026-07-24T13:57:24.000Z","lang":"en","categories":["concurrency","context","hardware","models","performance","quantization","runtime"],"models":["gemma-4-26b-a4b-it","gemma-4-31b","gemma-4-26b-a4b"],"posts":[{"id":"2080653579430539559","text":"Run Updated Gemma 4 26B A4B QAT (MoE) with Vision at 25 tokens/sec and massive 120k context window on a single RTX 4060 (8 GB VRAM + 16 GB RAM Only!!)\n\nYesterday I pushed Gemma 4 26B A4B QAT to 250k context on a single RTX 4060 using nothing but Q8 KV cache and optimized -b and -ub flags for higher prefill throughput.\n\nToday I stacked Multi Token Prediction (MTP) self speculative decoding AND the vision projector (mmproj) on top of that same card, same batch size optimization, same $250 GPU and pushed it until it broke, then found the fix.\n\nAll text only runs consist of a 28k prompt. vision runs consist of 28k text prompt + an image.\n\n# 1. MTP alone. near free decode speed, no catch\n\nMTP draft assistant is a separate small model (MTP heads are backed into the main model itself for the qwen 3.5+ models but its a separate small model for gemma 4 series), 240 MB gguf\n\n80k ctx: Prefill 510 t/s | Decode 29.5 t/s\n120k ctx: Prefill 433 t/s | Decode 29 t/s\n180k ctx: Prefill 240 t/s | Decode 24.9 t/s\n250k ctx: Prefill 63 t/s | Decode 13 t/s\n\nllama.cpp flags:\nm gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf-c 180000 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0\n\n# 2. Add vision on top. the tax you actually pay\n\nthe vision projector gguf is about 1.1 GBs\n\n80k ctx: Prefill 360 t/s | Decode 25.4 t/s\n120k ctx: Prefill 230 t/s | Decode 23.8 t/s\n180k ctx (Q8 KV): Prefill 75 t/s | Decode 12.5 t/s - cliff\n\nflags: \n-m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf --spec-type draft-mtp -md mtp-gemma-4-26B-A4B-it.gguf -c 80000 --port 8080 -b 1024 -ub 1024 --spec-draft-n-max 6 --spec-draft-p-min 0.7 -ctk q8_0 -ctv q8_0 --mmproj mmproj-F16.gguf\n\n# 3. The fix if you want to run vision over 120k context: swap Q8 KV for Q4 KV past 120k\n\nStack MTP + vision + Q8 KV past 120k context and you hit a wall. draft model overhead plus KV pressure tanks everything. Drop to Q4 KV and the wall disappears:\n\n180k ctx (Q4 KV): Prefill 220 t/s | Decode 25.5 t/s\n\n-ctk q4_0 -ctv q4_0 --mmproj mmproj-F16.gguf (rest same as above)\n\nBottom line: MTP gives you a near free +20-30% decode boost up to 120k context. Past that, it's fighting your VRAM, not helping and if vision is loaded too, Q4 KV isn't optional past 120k, it's mandatory. 30% boost is model and card specific, MTP boosted decode 2x for gemma 4 31b on a single rtx 4090.\n\nSame 8GB card. Same $250 GPU. Multimodal, speculative decoding, 180k usable context, zero upgrades.\n\nYou gotta try this if you have a single NVIDIA RTX 3050, 3060, 3070, 4050, 4060, 5050 or 5060. You can try it with a 6 GB VRAM card as well but you will have to lower the context window.\n\nHugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below.\n\nWhich models are you running on your 6/8/12GB cards with MTP?","created_at":"2026-07-24T13:57:24.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080653067582918656/vid/avc1/1920x1080/iD-vje85ZkAjYKtp.mp4?tag=29"}],"metrics":{"likes":29,"retweets":3,"replies":7,"views":1239},"quoted_tweet_id":"2080307789566722276"},{"id":"2080657784845283335","text":"No MTP No Vision (Baseline) https://t.co/bftGOMWfQ1","created_at":"2026-07-24T14:14:07.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN_7OSaakAASWN2.jpg:orig"}],"metrics":{"likes":2,"retweets":1,"replies":0,"views":55},"quoted_tweet_id":null}]},{"root_id":"2080650268929126529","url":"https://x.com/TheAhmadOsman/status/2080650268929126529","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-24T13:44:15.000Z","lang":"en","categories":["local","models","performance"],"models":[],"posts":[{"id":"2080650268929126529","text":"Opensource AI should remain usable, understandable, reproducible, locally deployable, economically viable, and community-governed even if today's dominant labs, foreign labs, hardware vendors, cloud platforms, or open-weight model providers change direction or disappear.\n\nIf intelligence becomes something people can only rent from a few closed institutions, the public does not just lose software freedom. It loses operational freedom.\n\nAI is a civilizational infrastructure for work, education, science, software, creativity, public services, and national capacity.\n\nThis civilizational infrastructure must not become rented access through closed APIs, remote platforms, shifting terms, opaque moderation, and prices set by a handful of companies.\n\nThe ability to study, build, repair, deploy, audit, adapt, teach, preserve, and run intelligence systems without asking permission is of EXISTENTIAL importance.\n\nWith OpenAI, Anthropic, and a handful of other players controlling the models, this civilizational infrastructure risks becoming a subscription economy for cognition.\n\nThe US should not fall behind on the freedom to run, inspect, modify, benchmark, teach, and preserve intelligence infrastructure. The practical posture is American capacity with global open standards.","created_at":"2026-07-24T13:44:15.000Z","media":[],"metrics":{"likes":69,"retweets":6,"replies":15,"views":36107},"quoted_tweet_id":"2080109289260159484"}]},{"root_id":"2080650105288634480","url":"https://x.com/jun_song/status/2080650105288634480","author":{"username":"jun_song","display_name":"Jun Song"},"created_at":"2026-07-24T13:43:36.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080650105288634480","text":"Jensen Huang’s very first post on X was about the importance of open-weight models.\n\nOpen source will win.","created_at":"2026-07-24T13:43:36.000Z","media":[],"metrics":{"likes":68,"retweets":3,"replies":7,"views":3536},"quoted_tweet_id":"2080643682408321103"}]},{"root_id":"2080649400947200237","url":"https://x.com/mr_r0b0t/status/2080649400947200237","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-24T13:40:48.000Z","lang":"en","categories":["context","models","performance"],"models":["qwen3-4b"],"posts":[{"id":"2080649400947200237","text":"Impressive research on context management 👀\n“Average gains of +7.2 / +3.2 / +6.4 across seven long-context benchmarks on Qwen3-4B / 8B / 30B-A3B, beating two prior RL training sets. Despite 64K training the gains hold to 512K+ tokens”","created_at":"2026-07-24T13:40:48.000Z","media":[],"metrics":{"likes":3,"retweets":0,"replies":1,"views":443},"quoted_tweet_id":"2080639139465015395"}]},{"root_id":"2080649294655127692","url":"https://x.com/aijoey/status/2080649294655127692","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-24T13:40:23.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080649294655127692","text":"Jensen blessed X with a policy letter, not a product drop. \n \n The thesis: America doesn’t win AI by one closed scoreboard. It wins by diffusion, factories, clinics, farms, main street, on models you can download, inspect, and run yourself. \n \n Open weights cut cost, force competition, let orgs own their stack, and put more defenders on the field. Risks are real; prohibition is worse. \n \nFrontier closed + frontier open. Build both.","created_at":"2026-07-24T13:40:23.000Z","media":[],"metrics":{"likes":8,"retweets":2,"replies":1,"views":335},"quoted_tweet_id":"2080643682408321103"}]},{"root_id":"2080643318325617033","url":"https://x.com/DerekColley_/status/2080643318325617033","author":{"username":"DerekColley_","display_name":"Derek Colley"},"created_at":"2026-07-24T13:16:38.000Z","lang":"en","categories":["concurrency"],"models":[],"posts":[{"id":"2080643318325617033","text":"Yesterday I attended the TS/AI conference in London, sponsored by @mastra @CopilotKit @sentry @elastic \n\nThe event was really well run, and the speakers were inspirational - thanks guys! 🙏\n\nSo much so, that I ignited an old side-quest...\n- to build an interactive voice ui for STT, TTS, VAD\n- multi session\n- queue/steer messages into Mastra AgentController\n\n📣 Code coming soon...","created_at":"2026-07-24T13:16:38.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN_tkFeXwAAXFmd.jpg:orig"}],"metrics":{"likes":7,"retweets":1,"replies":2,"views":136},"quoted_tweet_id":null},{"id":"2080644356235510116","text":"@mastra @CopilotKit @sentry @elastic Example of a queued message, waiting for the harness/agent controller - will be injected at the next turn https://t.co/f1siu02pcf","created_at":"2026-07-24T13:20:45.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN_u5WhXcAA1a5o.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":0,"views":13},"quoted_tweet_id":null},{"id":"2080672642357403885","text":"@mastra @CopilotKit @sentry @elastic as promised: https://t.co/uty8lZs9Vl","created_at":"2026-07-24T15:13:09.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":5},"quoted_tweet_id":null}]},{"root_id":"2080603956103750119","url":"https://x.com/superalesha/status/2080603956103750119","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-24T10:40:13.000Z","lang":"en","categories":["performance"],"models":["laguna-s-2-1"],"posts":[{"id":"2080603956103750119","text":"my benchmark scripts for Ling-3.0-flash vs Laguna S2.1 are ready","created_at":"2026-07-24T10:40:13.000Z","media":[],"metrics":{"likes":20,"retweets":0,"replies":1,"views":1651},"quoted_tweet_id":"2080554215144059027"}]},{"root_id":"2080580719638671454","url":"https://x.com/0xSero/status/2080580719638671454","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-24T09:07:53.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2080580719638671454","text":"If you've done work in OSS AI, or Local AI and want to have you and your company on\n\nState of Local AI - 2026\n\nGet in touch, you'll be on as a coauthor with the companies and people who make this all possible. \n\nWe moved release date to August 7th, we have 6 contributors so far https://t.co/s0J5Mzisn4","created_at":"2026-07-24T09:07:53.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080568305543196672/vid/avc1/2716x1664/uOvEI8-rOWF-H5Kp.mp4?tag=29"}],"metrics":{"likes":212,"retweets":11,"replies":37,"views":12890},"quoted_tweet_id":null}]},{"root_id":"2080562131938726132","url":"https://x.com/superalesha/status/2080562131938726132","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-24T07:54:01.000Z","lang":"en","categories":["agents","hardware","quantization","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2080562131938726132","text":"I deleted half the experts from Laguna S 2.1 INT4. the first version fit on 2x3090, then its reasoning completely fell apart.\n\nLaguna S 2.1 INT4 has 256 routed experts in each of 47 sparse layers. i wanted to keep 128 and shrink it from 66.96 to 37.18 GiB without requantizing the tensors that survived. \n\nfirst i used routing mass, basically keep the experts that receive the most router weight on my calibration corpus.\n\nthe surgery itself worked. 36,771 payload hashes matched the source, the checkpoint loaded at 18.7 GiB per GPU, vLLM served it on two 3090s. then i ran the frozen coding, tool and reasoning set.\n\nfull INT4: 64/75\nrouting-mass50: 45/75\n\nreasonmath went 12/15 -> 0/15. 13 of 15 generations hit the 16,384 token ceiling and completion tokens went from 8,509 to 213,262. the small token limit wasnt the problem after all, pruning had sent the reasoning into a loop.\n\nso i built the actual REAP profiler against the packed INT4 kernels. routing mass only counts popularity. REAP also looks at the size of each expert output, which is a lot closer to asking whether that expert changed the result.\n\ni profiled 3.2M rendered tokens across all 47 layers. compared with the failed mass selection, REAP replaced a median 46 of 128 experts per layer.\n\nREAP50 scored 56/75. reasonmath came back to 9/15, token-limit failures fell from 13 to 1 and completion tokens dropped to 19,382. much better. data extraction still got worse though, 8/15 versus 12/15 on full INT4.\n\nthen i kept all 256 experts and tried the full Q2 model. 34.71 GiB, so it also fits two 3090s.\n\nthe small matrix was already 63/75 for Q2 versus 56/75 for REAP50. on the complete coding suites:\n\nHumanEval+: 138/164 Q2, 119/164 REAP50\nMBPP+: 284/378 Q2, 247/378 REAP50\n\npaired MBPP+ made it even clearer. Q2 uniquely passed 56 tasks, REAP passed 19.\n\nREAP did its job. it rescued a terrible mass-pruning run and gave me a usable 37.18 GiB packed INT4 checkpoint for vLLM on two 3090s. it just didnt beat keeping every expert and squeezing the bits harder.\n\ni wanted REAP to win. it didnt.\n\ni started the full BFCL tool and agent run too, but agentic-memory was projecting another 20 to 50 hours after the coding result was clear. i stopped it and kept those rows as partial, not a score.\n\nQ2 used GGUF through llamacpp while REAP50 used packed INT4 through vLLM. \nsame prompts, templates, samplers, just different engines. not the cleanest a/b test, but overall pretty telling\n\nim really fascinated by the topic of pruning experts. i believe a lot of good things can come out of this.","created_at":"2026-07-24T07:54:01.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN-itEvWUAA6UU1.jpg:orig"}],"metrics":{"likes":3,"retweets":0,"replies":1,"views":266},"quoted_tweet_id":null}]},{"root_id":"2080490754053455941","url":"https://x.com/TheAhmadOsman/status/2080490754053455941","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-24T03:10:24.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2080490754053455941","text":"Excellent points for people who run models locally https://t.co/968AiqdhUU","created_at":"2026-07-24T03:10:24.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN9jOrpWUAAEi2a.png:orig"}],"metrics":{"likes":579,"retweets":61,"replies":22,"views":40776},"quoted_tweet_id":"2074872546692940059"}]},{"root_id":"2080481180915257656","url":"https://x.com/aijoey/status/2080481180915257656","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-24T02:32:21.000Z","lang":"en","categories":["concurrency","hardware","models","performance"],"models":[],"posts":[{"id":"2080481180915257656","text":"Two DGX Sparks , one QSFP cable between them, DeepSeek split across both. Six concurrent chats, 100+ tok/s aggregate, 1M token context. Nothing leaves the room. This is what it looks like behind the scenes. https://t.co/cjweZ25M6k","created_at":"2026-07-24T02:32:21.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080480823489241088/vid/avc1/1920x1080/VotzzdXfFhn_mvG-.mp4?tag=29"}],"metrics":{"likes":11,"retweets":0,"replies":5,"views":709},"quoted_tweet_id":null}]},{"root_id":"2080461343724114146","url":"https://x.com/Blackwellboy/status/2080461343724114146","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-24T01:13:32.000Z","lang":"en","categories":["quantization"],"models":[],"posts":[{"id":"2080461343724114146","text":"on this distribution you're really comparing 8.5B active of mostly code expertise against a dense 27B generalist. Even ternary quantized ..bonsai is degrading knowledge it actually has..quantization can't create knowledge laguna never trained in. Its worst tier being naturalistic prose says the same thing.. matches my testing exactly, it's why I only route Laguna to code and agent work currently","created_at":"2026-07-24T01:13:32.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":2,"views":121},"quoted_tweet_id":null}]},{"root_id":"2080451992670085224","url":"https://x.com/Blackwellboy/status/2080451992670085224","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-24T00:36:22.000Z","lang":"en","categories":["agents","concurrency","context","hardware","models","performance","quantization","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2080451992670085224","text":"Since posting this, my replies filled up with the same 6 complaints about Laguna S 2.1 loops, brutal cold starts, agent crashes, DFlash weirdness, memory pressure, long-context pain. Most people are hitting problems that already have fixes. Here's every issue from my testing: what's solved, what's workaround-able, and what's still genuinely broken 🧵➡️\n\n1️⃣ \"It gets stuck in endless reasoning loops\"\nMost-reported issue, especially on quantized builds in long agentic sessions. What helps: for benchmark/short-task work, disabling thinking (enable_thinking: false) sidesteps it entirely — that's how I ran my whole suite, deliberately, after early community loop reports. For agent sessions where you want reasoning on, there's no full config fix I know of yet — this one's on Poolside, and they know about it and hearing they will be shipping fix soon. \n\n2️⃣ \"Cold starts are horrific\"\nConfirmed and measured: ~10 min weight load, ~15 min first start when FlashInfer JIT + CUDA graph capture are included. No flag fixes this today. Workaround is architectural: treat it as a boot-and-leave lane, not on-demand. Persist your JIT cache across restarts (mount it in containers) so you only pay compilation once. Load-format experiments (InstantTensor-style) are on my list unverified so far, so I won't claim them.\n\n3️⃣ \"It breaks constantly connected to my agent framework\"\nThe one I can't fully validate my Laguna testing has been bench-shaped, not week-long production agent traffic, so honest flag on that. But two things from adjacent experience: if the breakage looks like garbled output under agent traffic, that pattern on other DFlash/spec-decode models traces to greedy draft sampling draft_sample_method: probabilistic + FlashInfer sampler + modest K fixed it on my DeepSeek lane. And verify your draft/target checkpoint revisions actually match a mismatched DFlash pair degrades silently. Also: never min_p under speculative decoding.\n\n4️⃣ \"DFlash acts weird at higher concurrency\"\nReal, measured, and it's a trade you can control. CUDA graph memory scales with seqs×(K+1), so config choices swing behavior hard: my sweep's K7/seqs=32 gives the best c=4 aggregate (61.7 tok/s) but low-seqs configs that win single-stream degrade under concurrency. There is no one right config there's a single-user profile and a multi-stream profile, and you should run the one matching your traffic (mine are K7/s8 and K7/s32). Also know DFlash is content-skewed regardless: ~45 tok/s on code, ~18 on free prose. That's not a bug, it's what draftability is.\n\n5️⃣ \"Memory pressure needs full pool plus swap\"\nTwo measured fixes. First: pin your KV cache (--kv-cache-memory-bytes, I use 12GiB → 327K KV tokens) instead of letting utilization-based allocation grab everything this freed ~28GB on my box versus default behavior. Second: drop max-num-seqs if you're single-user; seqs=32 vs 4 is a large graph-memory difference. Hard limit to respect: weights are 67GB and a serving stack holds ~99GB total two full instances do NOT co-reside on one 128GB Spark. I proved that one by accident this week.\n\n6️⃣ \"Long context gets stuck / is painful\"\nSplit this in two. Raw capability: verified working needle retrieval passes at 100K and 209K real cold tokens on my box (people hitting HTTP 400 near \"200K\": check your client's token math, that was my bug too). Latency: the real cost is cold prefill 45.6s TTFT at 100K, 133s at 209K, then ~14-19 tok/s decode. Anyone showing you fast long-context numbers without cold TTFT is showing you a warm cache. Prefix caching makes multi-turn agent sessions at depth livable first hit is expensive, the rest reuse it. Budget for the first hit.\n\nStill genuinely open: the reasoning loops (Poolside's court), agent-framework stability at scale (needs real soak testing, not benches on my list), and cold-start engineering. When it's running and routed to what it's good at code, tools, structured agent work on a private box it stays the most impressive single-Spark model I've used. The pain points are real; most have workarounds; none of them are mysteries.\n\nAll figures: rev 0761412, vLLM 0.25.1, FLASHINFER_CUTLASS native NVFP4, single GB10. Ask below if you want the exact flags for any of these. 👇","created_at":"2026-07-24T00:36:22.000Z","media":[],"metrics":{"likes":69,"retweets":5,"replies":12,"views":8380},"quoted_tweet_id":"2080309756317577509"},{"id":"2080464627297894885","text":"@eisokant little bit of community feedback for you guys :) also see original post. Hopefully it helps.","created_at":"2026-07-24T01:26:34.000Z","media":[],"metrics":{"likes":8,"retweets":0,"replies":2,"views":294},"quoted_tweet_id":null}]},{"root_id":"2080406053020127650","url":"https://x.com/MiaAI_lab/status/2080406053020127650","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-23T21:33:49.000Z","lang":"en","categories":["hardware","models","performance"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080406053020127650","text":"GLM-5.2 hits 22 tok/s on 3× DGX Sparks https://t.co/oUboKO5NJM","created_at":"2026-07-23T21:33:49.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN8VEqRXIAM8EQw.jpg:orig"}],"metrics":{"likes":146,"retweets":0,"replies":17,"views":6531},"quoted_tweet_id":null}]},{"root_id":"2080370423506587765","url":"https://x.com/mr_r0b0t/status/2080370423506587765","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-23T19:12:15.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080370423506587765","text":"How they gonna release a model this size and not make it open weights 😔","created_at":"2026-07-23T19:12:15.000Z","media":[],"metrics":{"likes":84,"retweets":1,"replies":8,"views":7419},"quoted_tweet_id":"2080351022028095681"}]},{"root_id":"2080346077442793871","url":"https://x.com/superalesha/status/2080346077442793871","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-23T17:35:30.000Z","lang":"en","categories":["hardware","models","performance"],"models":["qwen3-6-35b-a3b"],"posts":[{"id":"2080346077442793871","text":"half the experts. 53.7% of the file. same 146/164 on HumanEval+.\nI ran Qwen3.6 35B A3B coding-reap50 Q4 against the full Q4. \n11.37 GB vs 21.17 GB, both at 89.0%, fits one 3090\n\nthe size-matched shortcut, full Q2 at 12.94 GB, went to 121/164 😱 deleting half the experts cost nothing. squeezing the bits cost 15 points 👇","created_at":"2026-07-23T17:35:30.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN60j2faIAABj-r.jpg:orig"}],"metrics":{"likes":8,"retweets":1,"replies":6,"views":723},"quoted_tweet_id":null},{"id":"2080346081158963646","text":"speed was mixed.\npp512 +25.7%, TTFT at 8K -15.4%, but decode barely moved, 141.9 -> 144.2 tok/s.\nhalf the stored experts helped weight-heavy prefill, not decode. https://t.co/voYNquv7wH","created_at":"2026-07-23T17:35:31.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN61BIFbAAAUp-n.jpg:orig"}],"metrics":{"likes":1,"retweets":0,"replies":1,"views":76},"quoted_tweet_id":null},{"id":"2080347152241590423","text":"so im not calling this \"same quality everywhere\". GPQA direct was 16/50 vs 19/50 and that gap wasnt significant. kept every generation, exact argv and the pinned EvalPlus image.\n\nmodel: https://t.co/UFYihEpSAF","created_at":"2026-07-23T17:39:46.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":112},"quoted_tweet_id":null}]},{"root_id":"2080343905636159880","url":"https://x.com/TheAhmadOsman/status/2080343905636159880","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-23T17:26:52.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080343905636159880","text":"Data, evals, GPUs\n\nStudy any combination of those 3 and you’re gonna make great $$ in the next 5 years","created_at":"2026-07-23T17:26:52.000Z","media":[],"metrics":{"likes":663,"retweets":47,"replies":27,"views":26966},"quoted_tweet_id":null}]},{"root_id":"2080322016410780113","url":"https://x.com/MiaAI_lab/status/2080322016410780113","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-23T15:59:53.000Z","lang":"en","categories":["concurrency","models","performance","runtime"],"models":[],"posts":[{"id":"2080322016410780113","text":"Big update for sparkDash v1.3.0 ⚡️\n\n✨ LLM Prompt Showcase — full-page multi-terminal streaming demo (up to 32 concurrent prompts) from the LLM panel.\n\n✨ Aggregate server tok/s during runs, copy one/all terminals, collapsible reasoning, thinking-flag adapter.\n\n✨ Shared streaming path for Showcase + DecodeBench; mutual exclusion between the two.\n\n✨ Misc: UI compact mode, additional vLLM tiles, metrics history store, copy bench results to clipboard and more.\n\nGet it here\nhttps://t.co/8H3h5wjPOh","created_at":"2026-07-23T15:59:53.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080318657247248384/vid/avc1/3840x2160/a3KhSTad20aCEFzm.mp4?tag=29"}],"metrics":{"likes":30,"retweets":3,"replies":4,"views":1600},"quoted_tweet_id":null}]},{"root_id":"2080321061954933212","url":"https://x.com/StefanMaier/status/2080321061954933212","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-23T15:56:06.000Z","lang":"en","categories":["hardware","local","models"],"models":[],"posts":[{"id":"2080321061954933212","text":"Newest @NousResearch Hermes / local LLM / AI project on my @NVIDIAAI DGX Sparks:\n\nCreating a LLM / AI learning curricullum followed by a curated version for a audio book. Audio book is also already finished by tts in Hermes 🫡\n\nNow I have Full learning on the go 😀 https://t.co/a9kVab1Zrw","created_at":"2026-07-23T15:56:06.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN7I-k6XkAEzXkY.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HN7I-k9XcAA9r0E.jpg:orig"}],"metrics":{"likes":7,"retweets":2,"replies":3,"views":260},"quoted_tweet_id":null}]},{"root_id":"2080307789566722276","url":"https://x.com/analogalok/status/2080307789566722276","author":{"username":"analogalok","display_name":"Alok"},"created_at":"2026-07-23T15:03:21.000Z","lang":"en","categories":["agents","concurrency","context","hardware","models","performance","quantization","runtime"],"models":["gemma-4-26b-a4b-it","gemma-4-31b","gemma-4-26b-a4b"],"posts":[{"id":"2080307789566722276","text":"I just crammed the updated Gemma 4 26B A4B QAT (MoE) with 180k context into an 8GB RTX 4060 (8 GB VRAM + 16 GB RAM only!!) and optimized the batch size. 23 tokens/sec decode, 300 tokens/sec prefill \n\nYesterday I showed you a Gemma 4 31B dense model running flawlessly on an RTX 4090. Today, we're breaking the VRAM bank on a budget card using Unsloth’s new Gemma 4 26B (A4B) QAT quants. \n\nFollowing Google’s chat template update that boosted agentic benchmarks by +10%, I pushed this model to its absolute limits. Here is how you squeeze 250k context out of 8GB of VRAM.\n\n# The Setup & The Optimization\n\n- Hardware: Nvidia RTX 4060 (8GB VRAM) + 16GB System RAM\n- Environment: CUDA 13.0 build of llama.cpp\n- Model: gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf\n- Prompt: 28,000 tokens of prompt for each run\n\nIf you read my L2 cache breakdown (attached in replies), you know the 4060’s 24MB cache maxes out at `-b 1024 -ub 1024`. Push past that, and prefill crashes. I locked those flags in for every test below to ensure maximum GEMM throughput.\n\n# 1. The Raw Context Push (Unquantized KV Cache)\n\nFirst, I wanted to see how far pure 8GB VRAM + 16GB RAM could stretch without touching the KV cache:\n\n- 80k Context: Prefill 385 t/s | Decode 25.5 t/s\n- 120k Context: Prefill 270 t/s | Decode 24 t/s\n\nllama.cpp flags: \n.\\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 120000 --port 8080 -ub 1024 -b 1024\n\nWithout KV quantization, 120k is your hard ceiling. push past that prefill throughput drops off a cliff, making the model practically unusable for large agentic workloads.\n\n# 2. The Q8 KV Cache Lifeline\n\nTo survive 250k context on a budget card, you have to quantize the KV cache. I enabled 8 bit KV cache (`-ctk q8_0 -ctv q8_0`) and re ran:\n\n- 180k Context: Prefill 280 t/s | Decode 22.8 t/s\n- 250k Context: Prefill 115 t/s | Decode 20 t/s\n\nllama.cpp flags: \n.\\llama-server -m gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf -c 180000 --port 8080 -b 1024 -ub 1024 -ctk q8_0 -ctv q8_0\n\nResult: Q8 KV cache brings 250k context back from the dead. Decode speed stabilizes at a highly usable 20 t/s. You are trading a very small bit amount of reasoning precision for an extra 130,000 tokens of context window. \n\nif you own a single rtx 3050, 3060, 3070, 4050, 4060, 5050 or 5060, you must try this model and optimize your batch size for higher prefill.\n\nHugging Face links to the updated Unsloth's QAT quants and performance graph are in the replies below.\n\nWhat model are you running on your 6GB, 8GB or 12GB cards right now? Let's see your setups.","created_at":"2026-07-23T15:03:21.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080307421076176896/vid/avc1/1920x1080/tCLvIMS2a9KLl4Tj.mp4?tag=29"}],"metrics":{"likes":191,"retweets":23,"replies":16,"views":16592},"quoted_tweet_id":"2079212804809363681"},{"id":"2080659313518063691","text":"Multi Token Prediction (MTP) + Vision benchmarks are out! \n\nhttps://t.co/UCJAlblHl2","created_at":"2026-07-24T14:20:11.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":25},"quoted_tweet_id":"2080653579430539559"}]},{"root_id":"2080287553521868930","url":"https://x.com/sakurayukiai/status/2080287553521868930","author":{"username":"sakurayukiai","display_name":"Sakura Yuki"},"created_at":"2026-07-23T13:42:57.000Z","lang":"en","categories":["quantization"],"models":[],"posts":[{"id":"2080287553521868930","text":"@onusoz Antirez reproduced bad Italian on Q4 and the official API, so quantization gets an alibi. I’d check tokenizer fertility, then base vs instruct to separate weak pretraining from post-training collapse. https://t.co/NCZVp57CJl","created_at":"2026-07-23T13:42:57.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN6qf58WYAA1YhF.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":1,"views":163},"quoted_tweet_id":null}]},{"root_id":"2080282281088835611","url":"https://x.com/Blackwellboy/status/2080282281088835611","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-23T13:22:00.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2080282281088835611","text":"somewhere a man with no homelab just watches tv at night and falls asleep easily. we do not want his life. stop thinking about his life","created_at":"2026-07-23T13:22:00.000Z","media":[],"metrics":{"likes":6,"retweets":0,"replies":3,"views":138},"quoted_tweet_id":null}]},{"root_id":"2080280591988318434","url":"https://x.com/LottoLabs/status/2080280591988318434","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-23T13:15:17.000Z","lang":"en","categories":["hardware","performance"],"models":[],"posts":[{"id":"2080280591988318434","text":"Tobi with the DGX Station tps \n\nMight have to get those numbers on localmaxxing","created_at":"2026-07-23T13:15:17.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":195},"quoted_tweet_id":"2080270121390469469"}]},{"root_id":"2080271193093837304","url":"https://x.com/aijoey/status/2080271193093837304","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-23T12:37:56.000Z","lang":"en","categories":["hardware","local"],"models":["locateanything-3b"],"posts":[{"id":"2080271193093837304","text":"I’ve been trying to understand what “world models” could look like in a real industrial workflow.\n\nFor this demo, I started with one simulated ROV image of an intact subsea pipeline flange.\n\nNVIDIA Cosmos3-Edge, running locally on a single DGX Spark, generated a 10 second possible future where a bubble release begins at the joint and grows over time.\n\nThen I ran the generated frames through NVIDIA LocateAnything-3B. It found nothing at 0 or 2 seconds, then localized the bubble stream at 4, 6, 8, and 10 seconds.\n\nThat’s the part I find interesting: one model imagines a possible future, and another checks that future frame by frame and points the operator toward the emerging hazard.\n\nThis is a synthetic scenario, not an actual leak prediction or safety decision system. But it offers a glimpse of how open models could help people explore “what if?” scenarios before deciding what to inspect next.\n\n@NVIDIARobotics @NVIDIAAI","created_at":"2026-07-23T12:37:56.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080265872501772288/vid/avc1/1280x960/3xz0_0Yhke0L3_aB.mp4?tag=29"}],"metrics":{"likes":4,"retweets":1,"replies":1,"views":191},"quoted_tweet_id":null}]},{"root_id":"2080247818103796200","url":"https://x.com/LottoLabs/status/2080247818103796200","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-23T11:05:03.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2080247818103796200","text":"Seeing a lot less used 3090s in Canada 🥲","created_at":"2026-07-23T11:05:03.000Z","media":[],"metrics":{"likes":6,"retweets":0,"replies":6,"views":460},"quoted_tweet_id":null}]},{"root_id":"2080243186497179725","url":"https://x.com/antirez/status/2080243186497179725","author":{"username":"antirez","display_name":"antirez"},"created_at":"2026-07-23T10:46:39.000Z","lang":"en","categories":["models"],"models":["deepseek-v4-flash","laguna-s-2-1"],"posts":[{"id":"2080243186497179725","text":"So I tested Laguna S2.1 and the results are: it is a strong coding model for its size, in certain coding tasks it performs better than DeepSeek v4 Flash (but we are waiting for an update). It is NOT a general purpose LLM, very specialized for coding apparently.","created_at":"2026-07-23T10:46:39.000Z","media":[],"metrics":{"likes":243,"retweets":5,"replies":20,"views":17297},"quoted_tweet_id":null}]},{"root_id":"2080099061617578319","url":"https://x.com/mr_r0b0t/status/2080099061617578319","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-23T01:13:57.000Z","lang":"en","categories":["concurrency","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080099061617578319","text":"🚀 New benchmark results for Laguna-S 2.1 NVFP4 SM121 are in.\n\n⚡ 50.94 tok/s aggregate throughput at 8-way concurrency\n📈 Up to 1.27× higher throughput than the previous Laguna-S 2.1 NVFP4 weights\n\nConfigs, scripts, and raw results available in the repo ⬇️⬇️ https://t.co/MOsPkXhlrQ","created_at":"2026-07-23T01:13:57.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN3_EkhWgAEAOLM.jpg:orig"}],"metrics":{"likes":72,"retweets":5,"replies":6,"views":6217},"quoted_tweet_id":"2080095423973204062"}]},{"root_id":"2080072031689670762","url":"https://x.com/TheAhmadOsman/status/2080072031689670762","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-22T23:26:32.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2080072031689670762","text":"Fable 5 wasn't even available for majority of time Kimi K3 was being trained, any accusations of distillations CANNOT be true","created_at":"2026-07-22T23:26:32.000Z","media":[],"metrics":{"likes":765,"retweets":32,"replies":45,"views":31349},"quoted_tweet_id":null}]},{"root_id":"2080066129792303446","url":"https://x.com/MiaAI_lab/status/2080066129792303446","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-22T23:03:05.000Z","lang":"en","categories":["quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080066129792303446","text":"Yes seems like @poolsideai updated their Laguna-S-2.1 NVFP4. Delete the old weights and redownload. I'll re-test this version.","created_at":"2026-07-22T23:03:05.000Z","media":[],"metrics":{"likes":71,"retweets":7,"replies":11,"views":4895},"quoted_tweet_id":"2080063136216797678"}]},{"root_id":"2080063136216797678","url":"https://x.com/MiaAI_lab/status/2080063136216797678","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-22T22:51:12.000Z","lang":"en","categories":["quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2080063136216797678","text":"A new version of Laguna-S-2.1-NVFP4? \n\nhttps://t.co/rUZEm1CWgl https://t.co/35q8oFhkTc","created_at":"2026-07-22T22:51:12.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN3eQfYW4AAGWfy.jpg:orig"}],"metrics":{"likes":64,"retweets":2,"replies":10,"views":11313},"quoted_tweet_id":null}]},{"root_id":"2080058490341499158","url":"https://x.com/antirez/status/2080058490341499158","author":{"username":"antirez","display_name":"antirez"},"created_at":"2026-07-22T22:32:44.000Z","lang":"en","categories":["performance"],"models":["laguna-s-2-1"],"posts":[{"id":"2080058490341499158","text":"Still didn't attempt any quality testing, but let's say that GPT 5.6 Sol was able almost unassisted to write this Laguna S2.1 inference implementation following the other two models in DwarfStar, and it is very fast, 50 t/s generation, very fast prefill as well. https://t.co/O8nsbFEE7I","created_at":"2026-07-22T22:32:44.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080058185784688640/vid/avc1/3330x1860/eObalyROrFcnG07s.mp4?tag=29"}],"metrics":{"likes":281,"retweets":20,"replies":19,"views":23389},"quoted_tweet_id":null}]},{"root_id":"2080044212519923849","url":"https://x.com/aijoey/status/2080044212519923849","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-22T21:36:00.000Z","lang":"en","categories":["agents","hardware","models"],"models":[],"posts":[{"id":"2080044212519923849","text":"My last industrial AI demo showed Laguna solving a simulated problem inside an EV battery factory.\n\nThis time, I wanted to take it further and see what happens when the model becomes an agent that can investigate the problem for itself.\n\nQwen3.8 Max Preview runs through Hermes, watches live factory sensor data, checks maintenance records and tests different recovery options.\n\nAt the same time, NVIDIA Isaac Sim runs a live digital twin of the factory using OpenUSD and RTX rendering. NVIDIA WebRTC then streams that 3D environment directly into the browser control room.\n\nWhat I’m learning is that a digital twin can be much more than a 3D model.\n\nWhen you connect it to real operational data and an AI agent, it becomes a safe place to investigate failures and test possible solutions before changing anything in the real system.\n\nThe agent can find the problem and recommend a recovery.\n\nA human still makes the final call.\n\nThis all ran on one NVIDIA DGX Spark, the only cloud piece was the model itself. \n\ncc: @NVIDIAAI @NVIDIARobotics @NousResearch @nvidiaomniverse @Alibaba_Qwen","created_at":"2026-07-22T21:36:00.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080028281089187840/vid/avc1/1920x1080/2ncJ6h5hPuaCl_zI.mp4?tag=29"}],"metrics":{"likes":15,"retweets":1,"replies":3,"views":2368},"quoted_tweet_id":null}]},{"root_id":"2080040479337357524","url":"https://x.com/0xSero/status/2080040479337357524","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-22T21:21:10.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2080040479337357524","text":"They really did it! I'm so happy to see this being published. Now we have GLM-5.2 with vision. \n\nPutting those B300s to good use.\n\nThank you Baseten, I am porting this to the hybrid now.\nhttps://t.co/eh68Ox2NtY","created_at":"2026-07-22T21:21:10.000Z","media":[],"metrics":{"likes":1046,"retweets":83,"replies":40,"views":74438},"quoted_tweet_id":null}]},{"root_id":"2080035604624973852","url":"https://x.com/ivanfioravanti/status/2080035604624973852","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-22T21:01:47.000Z","lang":"en","categories":["agents","concurrency","context","local"],"models":[],"posts":[{"id":"2080035604624973852","text":"I respectfully disagree for several reasons.\n\nCalling a customer, whether free or paying, an idiot is simply wrong.\n\nOpenCode, like any other coding agent, clearly tries to preserve the prompt cache as much as possible. Otherwise, it would be painfully slow.\n\nThe “Stop Using OpenCode” article, which I believe Dax is referring to, focused on specific cases that can invalidate the cached prefix. These issues become much more visible and painful when using local models.\n\nModifying AGENTS.md is one example, as shown in this video. During this coding session, cache efficiency is extremely high. But the moment I modify AGENTS.md, boom: 62K tokens need to be processed again as new input.\n\nThe date change is another valid point raised in the article.\n\nI understand that both issues may sound trivial, and personally I've never been impacted by them, but their impact on the OpenCode end-user experience can be significant when using local models.\n\nI agree that both the feedback and the article were too direct and were probably written by the author out of frustration. However, they also contained constructive points that could help improve the product.\n\nI use OpenCode alongside several other tools, and I like it overall. But I really dislike this kind of reaction to user feedback, regardless of how that feedback is expressed.\n\nI would have preferred a response focused more on listening, understanding the problem, and improving the product.","created_at":"2026-07-22T21:01:47.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2080032995658924032/vid/avc1/1920x1080/if3lGrhODUjZkf5E.mp4?tag=29"}],"metrics":{"likes":140,"retweets":6,"replies":24,"views":21618},"quoted_tweet_id":"2079953961579102612"}]},{"root_id":"2080020279573541085","url":"https://x.com/RedHat_AI/status/2080020279573541085","author":{"username":"RedHat_AI","display_name":"Red Hat AI"},"created_at":"2026-07-22T20:00:54.000Z","lang":"en","categories":["hardware","models","runtime"],"models":["qwen3-5-397b-a17b"],"posts":[{"id":"2080020279573541085","text":"We just shipped DFlash speculator checkpoints for Qwen3.5-397B-A17B.\n\nOn math_reasoning: ~5 out of 7 draft tokens accepted on average. On code (HumanEval): ~4.5 out of 7.\n\nBoth checkpoints trained with the open source Speculators library from @vllm_project. Apache 2.0. Validated on NVIDIA H200.\n\nOne flag to enable in vLLM:\n\n --speculative-config '{\n \"model\": \"RedHatAI/Qwen3.5-397B-A17B-speculator.dflash\",\n \"num_speculative_tokens\": 7,\n \"method\": \"dflash\"\n }'\n\nCheck it out: https://t.co/k1QmN3djrv","created_at":"2026-07-22T20:00:54.000Z","media":[],"metrics":{"likes":55,"retweets":5,"replies":3,"views":2169},"quoted_tweet_id":null}]},{"root_id":"2079989602597634085","url":"https://x.com/ivanfioravanti/status/2079989602597634085","author":{"username":"ivanfioravanti","display_name":"Ivan Fioravanti ᯅ"},"created_at":"2026-07-22T17:59:00.000Z","lang":"en","categories":["agents","local"],"models":[],"posts":[{"id":"2079989602597634085","text":"The more I dig in the Local AI and Metal Kernel optimizations, the more I see that many people push performance without taking care of the fact that logits are drifting a lot compared to bf16/standard implementations.\n\nThis is not visibile with basic prompts/tests, but in evals and real usage, especially in agentic workflows you pay the price of this.\n\nWe all need to be more precise from this perspective.","created_at":"2026-07-22T17:59:00.000Z","media":[],"metrics":{"likes":18,"retweets":3,"replies":10,"views":878},"quoted_tweet_id":null}]},{"root_id":"2079989350960398445","url":"https://x.com/superalesha/status/2079989350960398445","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-22T17:58:00.000Z","lang":"en","categories":["hardware","models","performance","quantization","runtime"],"models":["qwen3-6-27b"],"posts":[{"id":"2079989350960398445","text":"there is a lock in your nvidia driver that makes your gpus 10x slower at talking to each other. its on by default. i patched it out.\n\ngpu-to-gpu latency went 14-18 us -> 1.5 us. bandwidth 5-6 -> 13.2 GB/s, 25.5 bidirectional. four used 3090s, no nvlink, no permission from anyone.\n\nthe hardware could always do this. p2p is disabled in software so the datacenter cards keep their moat.\n\nand now for the reality check: under actual vllm load it is +11-13%, not 10x.\nqwen3.5 122b awq, tp4: 65.5 -> 74.3 tok/s\nqwen3.6 27b int4 mtp, tp2: 98 -> 110-113 tok/s\n\nnot per request but aggregating throughput. anyone telling you about 2x is just full of it\n\n10x in synthetics and 13% in real life. $0 spent. because i can 😎","created_at":"2026-07-22T17:58:00.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN2bRwqbsAAUl9z.jpg:orig"}],"metrics":{"likes":52,"retweets":4,"replies":3,"views":1879},"quoted_tweet_id":null}]},{"root_id":"2079987476798615770","url":"https://x.com/mr_r0b0t/status/2079987476798615770","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-22T17:50:33.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079987476798615770","text":"🚨 Solar Open 2 is here—and it’s an open agent powerhouse.\n\n⚡ 250B MoE / 15B active\n🧠 1M-token context\n🛠️ Built for reasoning, coding & tool use\n📈 92.4 LiveCodeBench\n📈 70.4 SWE-bench Verified\n🌏 EN / KO / JA\nUpstage came to compete:\nhttps://t.co/jxa3ckFJ6N https://t.co/HTX1hPTiQd","created_at":"2026-07-22T17:50:33.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN2XskvXEAAiWY6.jpg:orig"}],"metrics":{"likes":10,"retweets":1,"replies":4,"views":458},"quoted_tweet_id":"2079949203615453414"}]},{"root_id":"2079978448622370888","url":"https://x.com/TeksEdge/status/2079978448622370888","author":{"username":"TeksEdge","display_name":"David Hendrickson"},"created_at":"2026-07-22T17:14:40.000Z","lang":"en","categories":["agents","context","hardware","local","models","performance","quantization"],"models":["gemma-4-12b"],"posts":[{"id":"2079978448622370888","text":"🎯 Local AI got a brand new 3B model with a novel architecture that beats Google’s Gemma 4 12B on major agentic and coding benchmarks.\n\nThis model was not a fine-tune or distillation of Qwen, Gemma, Llama, or another outside model\n\nThis is small enough to run on 4-8GB VRAM. Easily.\n💾 Full BF16 weights: about 8.4GB 📦 Q4 GGUF: about 2.6GB\n\nMeet Nanbeige4.2-3B, a small Apache-2.0 model designed for agents. 🤖\n\nVendor-reported results:\n🧑💻 SWE-bench Verified\n Nanbeige: 63.6\nGemma 4 12B: 44.2\n🛠️ SWE-bench Pro\n Nanbeige: 46.9\nGemma 4 12B: 21.9\n🖥️ Terminal-Bench 2.0\n Nanbeige: 44.1\nGemma 4 12B: 21.1\n\n🤖 Pinch-Bench V2\n Nanbeige: 74.7\nGemma 4 12B: 53.8\n🧠 GPQA Diamond\n Nanbeige: 87.4\nGemma 4 12B: 78.8\n\nHow can a 3B model punch this far above its weight?\n\n🔁 Looped Transformer architecture\nIt stores only 22 transformer layers and sends each token through those same layers twice, creating 44 effective layer executions without storing 44 separate layers.\n\nIn other words, it has a small memory footprint.\n\nSpecifications:\n🧠 3B non-embedding parameters\n📦 About 4B total parameters\n📚 262K context window\n🛠️ Tool calling in XML or JSON\n💭 Thinking and non-thinking modes\n🌐 English + Chinese\n🔓 Apache 2.0\n\nThe benchmark results come from Nanbeige, several agent evaluations depend on specific scaffolds, and Gemma 4 12B still wins on some instruction-following, recruitment and scientific-coding tests.\n\n🔗 HuggingFace Nanbeige/Nanbeige4.2-3B","created_at":"2026-07-22T17:14:40.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN2QZD4asAAjP_K.jpg:orig"}],"metrics":{"likes":546,"retweets":73,"replies":18,"views":54197},"quoted_tweet_id":"2079720420144816136"}]},{"root_id":"2079962217559253021","url":"https://x.com/LottoLabs/status/2079962217559253021","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-22T16:10:11.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079962217559253021","text":"@julien_c 🤣🤣 bbc editor heard huggingface and said we can’t have that published","created_at":"2026-07-22T16:10:11.000Z","media":[],"metrics":{"likes":4,"retweets":0,"replies":0,"views":276},"quoted_tweet_id":null}]},{"root_id":"2079959273363390901","url":"https://x.com/stevibe/status/2079959273363390901","author":{"username":"stevibe","display_name":"stevibe"},"created_at":"2026-07-22T15:58:29.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079959273363390901","text":"Image Models side-by-side\n\nAs Qwen has released Qwen Image 3.0 today, I have tested 4 prompts across 4 image models:\n\n> Qwen Image 3.0\n> GPT Image 2\n> Nano Banana 2\n> Seedream 5.0 Pro\n\nA thread 🧵 https://t.co/0GmUnKHyBk","created_at":"2026-07-22T15:58:29.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1_4HQaQAARyZD.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HN1_5Fwa8AAlVy0.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HN1_6G6bEAALUxJ.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HN1_7O_a0AAURiK.jpg:orig"}],"metrics":{"likes":171,"retweets":11,"replies":12,"views":19430},"quoted_tweet_id":null}]},{"root_id":"2079959095025508552","url":"https://x.com/TheZachMueller/status/2079959095025508552","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-22T15:57:46.000Z","lang":"en","categories":["runtime"],"models":[],"posts":[{"id":"2079959095025508552","text":"Thinky ( @thinkymachines ) model card is now out! Wanted to wait a few days for the SGLang & vLLM support to be fully in before publishing the numbers. \n\nOverall seems like a very fast and capable model, just be weary about it's modified Apache (one day maybe 🥺) https://t.co/VVCCN3hp9u","created_at":"2026-07-22T15:57:46.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1_qqzXwAABdk7.jpg:orig"}],"metrics":{"likes":22,"retweets":1,"replies":1,"views":970},"quoted_tweet_id":null}]},{"root_id":"2079945825204293948","url":"https://x.com/sudoingX/status/2079945825204293948","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-22T15:05:02.000Z","lang":"en","categories":["context","hardware","models","performance","quantization","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2079945825204293948","text":"running @poolsideai's laguna s 2.1 on a dgx spark hung my entire box. twice. so i spent a day tearing into why, and open sourced all of it so you skip straight to it running.\n\nfirst, the stack. docker is a dead end here, the public vllm image ships the wrong version and its kernels aren't built for the spark's gb10 chip. the path that works is native vllm 0.25.1 on cuda 13, plus a flashinfer nightly for the blackwell kernels.\n\nthen the three walls that each cost me an attempt. \n\n> one, you have to export CUTE_DSL_ARCH=sm_121a before you serve, or the nvfp4 moe kernels aren't compiled for gb10 and you get \"no kernel image for device,\" an error that tells you nothing. \n\n> two, you need python3.12-dev at runtime, not just at build, because triton jit-compiles a little c helper mid-inference and it needs python's headers, and enforce-eager does not save you. \n\n> three, the one that ate my box: on unified memory a big prefill working set spikes past the 121 gigs, the gpu is holding that memory so the oom killer can't reclaim it, and the whole machine locks, network and all. \n\nthe fix is small and non obvious, cap the prefill chunk with max num batched tokens 2048 and keep gpu memory utilization at 0.70 or under, and the working set stays tiny no matter how deep the context runs.\n\nthe config that actually holds is 128k context with the dflash drafter, that's the daily driver sweet spot. need the full million tokens? run it without the drafter, it's deep safe that way. what you cannot do is both, 256k plus dflash wedges the gpu, the drafter's kernels jit-compile for a giant novel shape mid inference and hang the driver, a known vllm bug, and poolside's fix for it is rejected by the nvfp4 quant. so you pick, the full context or the spec-decode speed.\n\nit's all in the repo, the install script, the crash safe serve wrapper, the benchmarks, the exact flags. apache-2.0, the next spark owner shouldn't lose the day i lost.\nhttps://t.co/lrWOlycv6I","created_at":"2026-07-22T15:05:02.000Z","media":[],"metrics":{"likes":54,"retweets":7,"replies":10,"views":2895},"quoted_tweet_id":null}]},{"root_id":"2079938767545413739","url":"https://x.com/superalesha/status/2079938767545413739","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-22T14:37:00.000Z","lang":"en","categories":["hardware","performance","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2079938767545413739","text":"a homie of mine runs the new Laguna S-2.1 on seven 3090s thru llama.cpp and gets 55 tok/s. i got around 100.. on four cards.\n\nsame gpus, same model, more than 2x difference. and its not magic, its parallelism.\n\nllama.cpp splits the model by layers by default, pipeline style. his gpus mostly stand in line waiting for each other. vllm with tensor parallel makes all cards chew every single layer together, so they actually work in parallel.\n\nmost people never touch this setting and lose half their throughput exactly like this. told him to switch to vllm, will see.\n\ncheatsheet 👇","created_at":"2026-07-22T14:37:00.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1PuVjawAEhXV-.jpg:orig"}],"metrics":{"likes":27,"retweets":1,"replies":3,"views":3200},"quoted_tweet_id":"2079780200837943724"}]},{"root_id":"2079936019369959750","url":"https://x.com/RedHat_AI/status/2079936019369959750","author":{"username":"RedHat_AI","display_name":"Red Hat AI"},"created_at":"2026-07-22T14:26:05.000Z","lang":"en","categories":["agents","context","models","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079936019369959750","text":"Poolside's Laguna S 2.1 is an open-weight MoE for agentic coding. 118B params, 8B active, 1M context. It competes with models many times its size and stays on a task for hours.\n\n@_soyr_ serves the FP8 version on @vllm_project and shows how you can too.\n\n@poolsideai https://t.co/elB465Oe5V","created_at":"2026-07-22T14:26:05.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079934363169423361/vid/avc1/1920x1080/K0XXIs2rNztB6JGw.mp4?tag=29"}],"metrics":{"likes":23,"retweets":4,"replies":3,"views":707},"quoted_tweet_id":null},{"id":"2079936020972261866","text":"poolside/Laguna-S-2.1-FP8: https://t.co/UjN8ZjSXUN\n\nFor more quantized models, check out Red Hat AI on Hugging Face: https://t.co/QRkOChCN1J","created_at":"2026-07-22T14:26:05.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":203},"quoted_tweet_id":null}]},{"root_id":"2079925889735790629","url":"https://x.com/MiaAI_lab/status/2079925889735790629","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-22T13:45:49.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079925889735790629","text":"@runsonai Qwen 3.6 is still the goal for small models","created_at":"2026-07-22T13:45:49.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":19},"quoted_tweet_id":null}]},{"root_id":"2079919750671688148","url":"https://x.com/LottoLabs/status/2079919750671688148","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-22T13:21:26.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2079919750671688148","text":"Gotta figure out how to get this on localmaxxing 🥹 only got 48GB vram","created_at":"2026-07-22T13:21:26.000Z","media":[],"metrics":{"likes":57,"retweets":0,"replies":13,"views":5751},"quoted_tweet_id":"2079612416967491952"}]},{"root_id":"2079910189755490428","url":"https://x.com/MiaAI_lab/status/2079910189755490428","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-22T12:43:26.000Z","lang":"en","categories":["agents","models","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079910189755490428","text":"Laguna-S-2.1 NVFP4 and Qwen 3.6 35B NVFP4 are neck-and-neck in tool eval bench for agentic workflows. Basically even.\n\nCoding tests next. https://t.co/Dq2Gbyl3FD","created_at":"2026-07-22T12:43:26.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1SQomW0AAOTlm.jpg:orig"}],"metrics":{"likes":72,"retweets":2,"replies":8,"views":2255},"quoted_tweet_id":null},{"id":"2079910536456671552","text":"https://t.co/skhI2h0mah","created_at":"2026-07-22T12:44:49.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1Tm3uXQAEbNko.jpg:orig"}],"metrics":{"likes":5,"retweets":0,"replies":2,"views":457},"quoted_tweet_id":null}]},{"root_id":"2079902964328632558","url":"https://x.com/mr_r0b0t/status/2079902964328632558","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-22T12:14:44.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079902964328632558","text":"Presenting Qwen3.8-Max Distillation 50K 🚀\nA curated dataset of 49,772 teacher-generated traces from qwen3.8-max-preview, prepared for supervised fine-tuning and off-policy knowledge distillation.\nUse responsibly 😛\nhttps://t.co/IU55a1bO2F https://t.co/FW27a0xKIV","created_at":"2026-07-22T12:14:44.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1MuUSW4AAM5uh.jpg:orig"}],"metrics":{"likes":76,"retweets":7,"replies":14,"views":3763},"quoted_tweet_id":"2079730799998353799"}]},{"root_id":"2079899321374548252","url":"https://x.com/stevibe/status/2079899321374548252","author":{"username":"stevibe","display_name":"stevibe"},"created_at":"2026-07-22T12:00:15.000Z","lang":"en","categories":["hardware","models","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079899321374548252","text":"Laguna S 2.1 running native NVFP4\n\nPicked Blackwell hardware for this side-by-side:\n🟢 DGX Spark | 19.44 tok/s | 162ms TTFT\n🔵 RTX PRO 6000 | 108.54 tok/s | 266ms TTFT\n🟣 4× RTX 5090 (TP4+EP4) | 145.81 tok/s | 521ms TTFT\n\n117.6B MoE (8.5B active), quantized to FP4, on your own box.","created_at":"2026-07-22T12:00:15.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079898900648079360/vid/avc1/2684x2160/z2un72fpFytya_dv.mp4?tag=29"}],"metrics":{"likes":97,"retweets":10,"replies":11,"views":8011},"quoted_tweet_id":null}]},{"root_id":"2079898870084223470","url":"https://x.com/superalesha/status/2079898870084223470","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-22T11:58:27.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079898870084223470","text":"Actually, it is not necessary for Open Source LLMs to win.\nIt is enough for them to keep pace so that closed-source labs cannot set any price they want.","created_at":"2026-07-22T11:58:27.000Z","media":[],"metrics":{"likes":2,"retweets":1,"replies":1,"views":166},"quoted_tweet_id":null}]},{"root_id":"2079897711197720900","url":"https://x.com/MiaAI_lab/status/2079897711197720900","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-22T11:53:51.000Z","lang":"en","categories":["hardware","models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079897711197720900","text":"GLM 5.2 running on 3x DGX Sparks https://t.co/FpeeUYhyjM","created_at":"2026-07-22T11:53:51.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN1H5oZXkAAdQcg.jpg:orig"}],"metrics":{"likes":83,"retweets":4,"replies":15,"views":3211},"quoted_tweet_id":null},{"id":"2080275409308909759","text":"Recipe published\n\nhttps://t.co/vqJrKX8xf0","created_at":"2026-07-23T12:54:41.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":161},"quoted_tweet_id":"2080274019035619769"}]},{"root_id":"2079890894287732822","url":"https://x.com/TheZachMueller/status/2079890894287732822","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-22T11:26:46.000Z","lang":"en","categories":["models","quantization"],"models":[],"posts":[{"id":"2079890894287732822","text":"Okay. Now that GLM is deployed and somewhat fast, time to…. Try and deploy DSv4 Flash on 8xH100’s and see how fast I can make it go. \n\nProbably DP4, EP8, TP2, + FP8 kv","created_at":"2026-07-22T11:26:46.000Z","media":[],"metrics":{"likes":9,"retweets":0,"replies":1,"views":1217},"quoted_tweet_id":null}]},{"root_id":"2079881678953349406","url":"https://x.com/sudoingX/status/2079881678953349406","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-22T10:50:09.000Z","lang":"en","categories":["agents","hardware","models","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079881678953349406","text":"poolside's laguna s 2.1 is live on my dgx spark rn! 118b moe, 8.5b active per token, and the whole thing runs on one box that fits on a desk.\n\nthe ram footprint: ~67 gigs of nvfp4 weights, and the full million token context adds only ~25 gigs of kv (hybrid attention, just 12 of 48 layers are full attn), so it all lands near 100 of 128 gigs with room to spare.\n\nspeed, honestly: single stream it's modest, ~19 tok/s, this box was never a sprinter. but with the dflash drafter it climbs to 25-30 on normal chat and 45 sustained on code, and under load it pushes ~140 tok/s across 16 streams. plenty to talk to in real time.\n\nand it is good to talk to, coherent and quick. but talking is the easy part. the real question with a poolside model is whether it can WORK. \n\nagentic capabilities next, let's see if it can actually build.","created_at":"2026-07-22T10:50:09.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HN04ctlacAANPSk.jpg:orig"}],"metrics":{"likes":90,"retweets":4,"replies":13,"views":7327},"quoted_tweet_id":"2079623151953310190"}]},{"root_id":"2079806664728064118","url":"https://x.com/superalesha/status/2079806664728064118","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-22T05:52:04.000Z","lang":"en","categories":["hardware","performance"],"models":[],"posts":[{"id":"2079806664728064118","text":"8.5b active params run perfectly on my 3090, the other 109b are just chilling in vram\n\nfield notes: ~850W at the wall under load. the DFlash draft model? tried it, only helps solo code gen, left it off.\n\nmost usable 100B+ this rig has ever run https://t.co/zWvXfmKBXV","created_at":"2026-07-22T05:52:04.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNz0FHXbUAIBjFn.jpg:orig"}],"metrics":{"likes":7,"retweets":0,"replies":2,"views":213},"quoted_tweet_id":null},{"id":"2079806668892958862","text":"Laguna ships with a 1.1B DFlash draft for speculative decoding. official numbers say 2.5-3.8x.\n\non my TP4 3090s exactly one workload wins: solo code gen, 99 -> 152 tok/s. random text and batches got slower.\n\nthe draft is a mind reader. it only pays off when your text is predictable.","created_at":"2026-07-22T05:52:05.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNz07rEaoAAAP6u.jpg:orig"}],"metrics":{"likes":3,"retweets":0,"replies":0,"views":185},"quoted_tweet_id":null}]},{"root_id":"2079806656897311068","url":"https://x.com/superalesha/status/2079806656897311068","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-22T05:52:02.000Z","lang":"en","categories":["concurrency","context","hardware","models","performance","quantization","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2079806656897311068","text":"Laguna S 2.1 runs on four used 3090s. full 262K context, no cuts. The best model for my setup.\n\n117B MoE, INT4, vLLM. 98 tok/s single request, 503 tok/s at 64 concurrent.\n\nhere is what it does on $6K of used hardware. 152 measurements. https://t.co/QpUpwjD3xI","created_at":"2026-07-22T05:52:02.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNzzvOWa8AA7kP-.jpg:orig"}],"metrics":{"likes":34,"retweets":4,"replies":12,"views":2376},"quoted_tweet_id":null},{"id":"2079807189641056751","text":"great work @poolsideai","created_at":"2026-07-22T05:54:09.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":167},"quoted_tweet_id":null},{"id":"2079884276108235025","text":"full benchmarks here:\nhttps://t.co/40ZKrR59jp","created_at":"2026-07-22T11:00:28.000Z","media":[],"metrics":{"likes":3,"retweets":0,"replies":0,"views":116},"quoted_tweet_id":null}]},{"root_id":"2079805367459139607","url":"https://x.com/NeoAIForecast/status/2079805367459139607","author":{"username":"NeoAIForecast","display_name":"Neo"},"created_at":"2026-07-22T05:46:55.000Z","lang":"en","categories":["hardware","local","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079805367459139607","text":"Laguna-S-2.1-NVFP4 running locally on the DGX Spark.\n\nOneshot with my standard prompt https://t.co/QqW0zFs4Sw","created_at":"2026-07-22T05:46:55.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079804258472579072/vid/avc1/1278x688/52h3W4UQljn-6a6B.mp4?tag=29"}],"metrics":{"likes":33,"retweets":0,"replies":1,"views":2390},"quoted_tweet_id":null}]},{"root_id":"2079777620464710008","url":"https://x.com/Blackwellboy/status/2079777620464710008","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-22T03:56:39.000Z","lang":"en","categories":["local","models"],"models":[],"posts":[{"id":"2079777620464710008","text":"OpenAI hacking hugging face sounds like centralised trying to take down open sourced. How convenient our main source of local models gets targeted.","created_at":"2026-07-22T03:56:39.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":1,"views":85},"quoted_tweet_id":null}]},{"root_id":"2079758318470565933","url":"https://x.com/bleysg/status/2079758318470565933","author":{"username":"bleysg","display_name":"Bleys Goodson"},"created_at":"2026-07-22T02:39:57.000Z","lang":"en","categories":["hardware","performance"],"models":[],"posts":[{"id":"2079758318470565933","text":"I was tempted to dismiss '14nm ≈ 4nm' as cope. Then I found the big number, and it isn't TFLOPS/W. It's FLOPs per sanctioned input: infinite. Zero advanced wafers, zero HBM, zero CoWoS. Orient Computing Core's DF1000 mints AI compute from inputs export controls barely touch.\n\nAs a chip, the dunks land: 0.33–0.45 TFLOPS/W is 14nm physics, ~3–4x the electricity of Blackwell-class per peak FLOP, power column missing from every slide. But the decision-relevant quantity is ΔFLOPs to China's aggregate.\n\n'Scale Huawei' hits a wall fast: Ascend output is capped by whichever runs out first, sanctioned scanner-hours or HBM. DF1000 dodges the wall entirely: 14nm-class wafers (no SAQP scanner tax, no overlap with Ascend's litho fleet, expandable with tooling China can still buy or build), commodity DRAM, domestic bonding. Roughly 100–150k wpm of 14nm-class capacity would match the whole advanced-node AI pipeline in throughput even at several-fold worse FLOPS per wafer: a buildout target, not a moonshot, against a mature base already in the millions of wpm.\n\nThe second win is allocation. Advanced silicon has comparative advantage in training (per-chip capability, interconnect, CANN); bandwidth-rich FLOP-poor stacks have it in decode-heavy inference on cheap watts. Every serving workload absorbed releases a rationed Ascend back to training. Score it partly in freed advanced silicon.\n\nThe watt cost has a ramp. Today it's cheap: this hardware burns 3–4x the electricity per FLOP, so it belongs on China's cheapest power (curtailed wind and solar, off-peak, interruptible supply), where the watts have almost no competing use. If a global power crunch arrives and even China's grid tightens, the cost bites: efficient chips claim the prime slots and these get pushed down the power-quality ladder. But pushed down, not scrapped: you can only retire a chip as fast as you can build its replacement, and replacement silicon stays rationed in every crunch scenario.\n\nWhat to watch: independently measured MFU (sets whether the effective gap is 3.7x or under 2x), stacked-DRAM thermals and yield (Wei Shaojun concedes the yield challenge himself), and whether wafer-level DRAM-on-logic bonding scales, because that's upstream learning for HiBL, LogicFolding, and the named successors: DF2000 in Q4 2026 targeting H200, DF3000 in 2027 targeting B300, where the disposable 14nm logic gets swapped and the near-memory system stays.\n\nTheir 4nm slogan is naughty spin, but the supply-curve intervention is a real capacity unlock for China's AI factories.","created_at":"2026-07-22T02:39:57.000Z","media":[],"metrics":{"likes":3,"retweets":0,"replies":0,"views":1550},"quoted_tweet_id":"2079730898791211010"}]},{"root_id":"2079752804323807647","url":"https://x.com/SpaceTimeViking/status/2079752804323807647","author":{"username":"SpaceTimeViking","display_name":"ÆON FORGE ✨"},"created_at":"2026-07-22T02:18:03.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2079752804323807647","text":"This concept art video was fully rendered on a local SINGLE @NVIDIAAI DGX Spark by the way.","created_at":"2026-07-22T02:18:03.000Z","media":[],"metrics":{"likes":47,"retweets":1,"replies":5,"views":3111},"quoted_tweet_id":"2079752537192685904"}]},{"root_id":"2079752537192685904","url":"https://x.com/SpaceTimeViking/status/2079752537192685904","author":{"username":"SpaceTimeViking","display_name":"ÆON FORGE ✨"},"created_at":"2026-07-22T02:16:59.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2079752537192685904","text":"A lot of you know me through my work on AI models and optimizing local AI for DGX Spark (among other things).\n\nWhat you probably don’t know is the real reason I’ve been building AI in the first place:\n\nTo build something I’ve had in the works for decades.\n\nTerraForge\nA preview https://t.co/JjYxuSYwtd","created_at":"2026-07-22T02:16:59.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079723568183549952/vid/avc1/832x480/xTEcQgxdIEhUpm_Z.mp4?tag=29"}],"metrics":{"likes":92,"retweets":5,"replies":11,"views":7905},"quoted_tweet_id":null}]},{"root_id":"2079745233240084988","url":"https://x.com/WescheNex1q/status/2079745233240084988","author":{"username":"WescheNex1q","display_name":"Wësche"},"created_at":"2026-07-22T01:47:58.000Z","lang":"en","categories":["agents","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079745233240084988","text":"Laguna-S-2.1-NVFP4 think ON vs OFF:\n\nON: Code 70→73, instructions 86→92\nPlanning 65→70. \nReliability on agentic fell hard (98.6→86.3). Latency 5.9s→7.8s.\n\nOverall: 86.5 off · 86.0 on. \n\nThinking is a dial, not an upgrade.\nON to write code. OFF to run agents.\n\nRecipe + board: https://t.co/w3XdoBror5\nhttps://t.co/Qk2CcWXN3q","created_at":"2026-07-22T01:47:58.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNy8mS4XsAAsGno.jpg:orig"}],"metrics":{"likes":59,"retweets":2,"replies":4,"views":3412},"quoted_tweet_id":null}]},{"root_id":"2079730799998353799","url":"https://x.com/mr_r0b0t/status/2079730799998353799","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-22T00:50:36.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079730799998353799","text":"Qwen3.8-Max-Preview distilling Qwen3.8-Max-Preview: hour 13 update\n32,987/50,000 traces produced\nZERO refusals 😛 https://t.co/C4wEpGIDaM","created_at":"2026-07-22T00:50:36.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNywJCmWwAAANE5.jpg:orig"}],"metrics":{"likes":49,"retweets":2,"replies":7,"views":1562},"quoted_tweet_id":null}]},{"root_id":"2079703415786012869","url":"https://x.com/0xSero/status/2079703415786012869","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-21T23:01:48.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079703415786012869","text":"We tinkered for months, doing everything we can to push the local frontier.\n\nFriday we will share our findings.","created_at":"2026-07-21T23:01:48.000Z","media":[],"metrics":{"likes":291,"retweets":4,"replies":10,"views":13249},"quoted_tweet_id":"2079702822887584221"}]},{"root_id":"2079703162647183499","url":"https://x.com/TheAhmadOsman/status/2079703162647183499","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-21T23:00:47.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079703162647183499","text":"BREAKING\n\nChinese opensource model “GLM 5.2” intercepts cyberattacks on a US corporation\n\nSource of attack identified as another US corporation “OpenAI” which serves closed proprietary models","created_at":"2026-07-21T23:00:47.000Z","media":[],"metrics":{"likes":815,"retweets":86,"replies":29,"views":27654},"quoted_tweet_id":null}]},{"root_id":"2079683537226330380","url":"https://x.com/superalesha/status/2079683537226330380","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-21T21:42:48.000Z","lang":"en","categories":["hardware","local","models"],"models":["laguna-s-2-1"],"posts":[{"id":"2079683537226330380","text":"this FPS gameplay was generated by a model running on 4 gaming GPUs in my apartment\n\nnot a cloud demo. not an API. my own 96GB rig\n\nLaguna-S-2.1 dropped, it fits, and I still can't believe what I'm looking at\n\nlocal video models just entered their GOAT era\n\nhttps://t.co/jb7r2inkt9","created_at":"2026-07-21T21:42:48.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079683089576660993/vid/avc1/1920x1080/YbBwAPnx-Exee8Fg.mp4?tag=29"}],"metrics":{"likes":15,"retweets":0,"replies":5,"views":2681},"quoted_tweet_id":"2078772825863626958"},{"id":"2079683539768094845","text":"actually couldn't pull off the one-shot. my prompt was too heavy and complex for it to handle all at once. had to tinker with laguna to fix it up.","created_at":"2026-07-21T21:42:49.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":114},"quoted_tweet_id":null},{"id":"2079684520044028236","text":"another interesting thing i noticed.\ni used the same prompt to test qwen3.8, fable 5, and kimi k3 -they all output a very similar style. like, literally identical. different implementations, but the style is so close. laguna is different though, and it stands out a lot. i dont know why, but i have two main guesses:\n1. its just too small and doesnt fully grasp the prompt details.\n2. it was trained totally differently, without the heavy distillation the others went through (from each other). i find that hard to believe.","created_at":"2026-07-21T21:46:42.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":1,"views":239},"quoted_tweet_id":null}]},{"root_id":"2079678459908063604","url":"https://x.com/aijoey/status/2079678459908063604","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-21T21:22:38.000Z","lang":"en","categories":["agents","hardware","local","performance"],"models":["laguna-s-2-1"],"posts":[{"id":"2079678459908063604","text":"Tired of seeing the same benchmarks around local models? I ran @poolsideai’s Laguna S 2.1 locally on a single NVIDIA DGX Spark.\n\nI gave it a simulated problem inside an electric vehicle battery factory.\n\nOne of the machines kept producing parts with bad torque readings. The robot’s power usage and vibration were slowly increasing, but not enough to trigger any alarms.\n\nMaintenance thought the robot was failing.\n\nProduction thought the factory was receiving defective parts.\n\nMeanwhile, more products were being rejected and the production line was starting to back up.\n\nLaguna reviewed the factory sensor data, worker notes, maintenance manuals, machine limits, quality records and the production schedule.\n\nIt ruled out the incoming parts and found that several small warning signs pointed to a worn spindle bearing.\n\nBefore changing anything, it tested a solution inside a digital factory simulation. It rerouted production to a backup machine, checked the results and then stopped to ask for human approval before continuing.\n\nThe results:\n\nOEE improved from 71.4% to 93.1%\n\nTorque rejects dropped from 8.4% to 1.1%\n\nThe production queue dropped from 18 packs to 5\n\nThe factory itself was simulated.\n\nThe model run, tool calls and investigation were real and happened locally on one DGX Spark.\n\nThis is not about letting AI control a factory by itself.\n\nIt is about giving people another tool to understand complicated problems faster, connect signals they might normally miss and test possible solutions before touching the real production line.\n\nReally interesting work from @poolsideai, @eisokant and @jasoncwarner.","created_at":"2026-07-21T21:22:38.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079676944371593216/vid/avc1/1920x1080/NNMo-Pe1fnX1cQBy.mp4?tag=29"}],"metrics":{"likes":82,"retweets":7,"replies":13,"views":6066},"quoted_tweet_id":null}]},{"root_id":"2079676950042349783","url":"https://x.com/mr_r0b0t/status/2079676950042349783","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-21T21:16:38.000Z","lang":"en","categories":["quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079676950042349783","text":"Taking the time to make sure @poolsideai Laguna-S-2.1-NVFP4 is fully optimized for SM121 ❤️ https://t.co/IU4gMMrLOx","created_at":"2026-07-21T21:16:38.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNx_KZKWwAE6NOB.jpg:orig"}],"metrics":{"likes":51,"retweets":1,"replies":1,"views":1746},"quoted_tweet_id":null}]},{"root_id":"2079667815892816214","url":"https://x.com/StefanMaier/status/2079667815892816214","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-21T20:40:20.000Z","lang":"en","categories":["hardware","local","models"],"models":[],"posts":[{"id":"2079667815892816214","text":"What I love about the @NVIDIAAI DGX Spark community: through tinkering, a bit of magic, and some love, we transform an underrated device with comparatively low bandwidth into an AI powerhouse—one that now even conjures \"Frontier\" models right into our homes or offices.\n\nLocal LLMs must (and will) win,\nOpen Source must (and will) win!","created_at":"2026-07-21T20:40:20.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNx22xZW4AANNlh.jpg:orig"}],"metrics":{"likes":29,"retweets":3,"replies":6,"views":1137},"quoted_tweet_id":null}]},{"root_id":"2079661740791574857","url":"https://x.com/Youssofal_/status/2079661740791574857","author":{"username":"Youssofal_","display_name":"Youssof Al Toukhi"},"created_at":"2026-07-21T20:16:11.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079661740791574857","text":"Initial thoughts: \n\nI don’t like it. \n\nOn open router is has this obnoxious hybrid reasoning built in where it determines your request doesn’t need reasoning tokens even when reasoning is set to max. \n\nWhen I send a complex math question it reasons.\n\nOn code it doesn’t, then outputs slop.\n\nIt may be a bug, I’ll test it locally.","created_at":"2026-07-21T20:16:11.000Z","media":[],"metrics":{"likes":47,"retweets":0,"replies":12,"views":11408},"quoted_tweet_id":"2079613777343848465"}]},{"root_id":"2079659526945636583","url":"https://x.com/sudoingX/status/2079659526945636583","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-21T20:07:24.000Z","lang":"en","categories":["context","hardware","models","performance","quantization"],"models":["laguna-s-2-1"],"posts":[{"id":"2079659526945636583","text":"they built this 118b moe model from scratch, its full 1million ctx, all of it on a single dgx spark that fits on my desk. the box has been waiting for a model shaped exactly like this.\n\npoolside just gave every spark owner laguna s 2.1: a 118b mixture of experts, 8.5b active per token, 1m context, open weights. not a llama or qwen tune, their own architecture and vocab, built from zero.\n\nwhy it's the spark's model specifically: moe only fires 8.5b of the 118b per token, so decode stays quick on the box's bandwidth where a dense 118b would crawl, dragging every weight through memory every single token. big memory to hold it, few active params to keep it moving, that is the spark thesis in one model.\n\nand the whole thing fits. the nvfp4 weights come in at ~67 gigs. i'd calculated ~65 before i ever loaded it, the napkin math held on metal. \n\nthe trick that makes 1m context fit: only 12 of the 48 layers run full attention, the rest slide, and the kv is fp8, so the million-token cache is ~26 gigs instead of the ~250 a normal 118b would demand. weights plus the full window lands near 100 gigs, on one 128gb spark, with real headroom left.\n\na frontier class 118b holding its entire million token context, on a box that fits on your desk. numbers soon.\n\nmodel: https://t.co/sILaslU11H","created_at":"2026-07-21T20:07:24.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNxstR5bIAAIK5b.jpg:orig"}],"metrics":{"likes":155,"retweets":14,"replies":18,"views":9858},"quoted_tweet_id":"2079623151953310190"}]},{"root_id":"2079642634184958076","url":"https://x.com/PrismML/status/2079642634184958076","author":{"username":"PrismML","display_name":"PrismML"},"created_at":"2026-07-21T19:00:16.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079642634184958076","text":"Try Ternary Bonsai 27B directly on Hugging Face , powered by @togethercompute \n\nhttps://t.co/yB7uIaHXpF https://t.co/PEDLAI2C6u","created_at":"2026-07-21T19:00:16.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079642037184503808/vid/avc1/3022x1734/2lTvZHuAtctxawuP.mp4?tag=29"}],"metrics":{"likes":57,"retweets":5,"replies":3,"views":6450},"quoted_tweet_id":null}]},{"root_id":"2079642410725118326","url":"https://x.com/MiaAI_lab/status/2079642410725118326","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-21T18:59:23.000Z","lang":"en","categories":["hardware"],"models":["laguna-s-2-1"],"posts":[{"id":"2079642410725118326","text":"Laguna S 2.1 up and running on DGX Spark. Now let's see what it can do, should be interesting!","created_at":"2026-07-21T18:59:23.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNxfk7-WEAAFIs3.jpg:orig"}],"metrics":{"likes":206,"retweets":3,"replies":29,"views":31866},"quoted_tweet_id":null}]},{"root_id":"2079628791136100566","url":"https://x.com/Sentdex/status/2079628791136100566","author":{"username":"Sentdex","display_name":"Harrison Kinsley"},"created_at":"2026-07-21T18:05:16.000Z","lang":"en","categories":["hardware","local","quantization"],"models":[],"posts":[{"id":"2079628791136100566","text":"118B A8B is comfy AF. You'll be able to FP8 this model on 128GB machines, great for various macs, sparks, gb10s...etc. \n\nI got to play with this model a bit in preview and it definitely feels insanely good for it's size. Will play more locally soon!","created_at":"2026-07-21T18:05:16.000Z","media":[],"metrics":{"likes":218,"retweets":12,"replies":17,"views":12158},"quoted_tweet_id":"2079613777343848465"}]},{"root_id":"2079623151953310190","url":"https://x.com/sudoingX/status/2079623151953310190","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-21T17:42:51.000Z","lang":"en","categories":["context","hardware","local","models"],"models":["deepseek-v4-pro","laguna-s-2-1"],"posts":[{"id":"2079623151953310190","text":"this is the drop the local ai crowd should be losing their minds over.\n\npoolside just dropped laguna s 2.1: 118b total parameters, only 8b active per token, a full 1m context window, open weights under a real open license, on huggingface today.\n\nlook at the chart. it lands at 71 on terminal-bench at 118b, sitting above deepseek v4 pro max at a trillion params, above inkling at 1.5 trillion, above nemotron 3 ultra. it's beating models ten times its size and losing only to kimi k3, which is 24 times bigger. that's the efficiency frontier, up and to the left, exactly where you want a model to sit.\n\nbut here's the part that made me sit up: it runs on a single dgx spark.\n\nand this is what nobody's saying loud enough. the dgx spark is the moe king. a dense 118b would crawl on it, the bandwidth chokes reading every weight each token. a moe with 8b active only ever reads 8b, so the spark's 128 gigs holds the whole model while generation stays fast. big brain, light footprint, the exact shape the spark was built to run.\n\nopen, frontier competitive, moe efficient, and it fits on a box on your desk. that's the whole thesis in one release: you don't need a datacenter, you need the right architecture on the right hardware. go grab the link below, weights are up.","created_at":"2026-07-21T17:42:51.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNxKZw5bMAA9R76.jpg:orig"}],"metrics":{"likes":1060,"retweets":110,"replies":62,"views":144983},"quoted_tweet_id":"2079613777343848465"}]},{"root_id":"2079619615756337274","url":"https://x.com/TheAhmadOsman/status/2079619615756337274","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-21T17:28:48.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079619615756337274","text":"OpenAI sent me an email to start advertising to ChatGPT users\n\nThe data they are collecting from your conversations will be used to serve ads to you\n\nYet another reason that Local and Opensource AI must win https://t.co/4DFqAOgMj4","created_at":"2026-07-21T17:28:48.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNxKwahW0AAxwDB.jpg:orig"}],"metrics":{"likes":22,"retweets":4,"replies":3,"views":739},"quoted_tweet_id":"2079406563370951064"}]},{"root_id":"2079617080643707331","url":"https://x.com/vllm_project/status/2079617080643707331","author":{"username":"vllm_project","display_name":"vLLM"},"created_at":"2026-07-21T17:18:44.000Z","lang":"en","categories":["agents","context","hardware","local","models","quantization","runtime"],"models":["laguna-s-2-1"],"posts":[{"id":"2079617080643707331","text":"🎉 Congrats to @poolsideai on Laguna S 2.1, a new open-weight model built for agentic coding and long-horizon work.\n\n🧠 118B sparse MoE, only 8B active per token, up to 1M context, thinking + no-thinking modes, OpenMDW-1.1\n\n🔁 Built to stay on task across long, multi-step runs: plan, call tools, check its work, recover, keep going\n\n🖥️ The official NVFP4 quant runs locally on a single @NVIDIAAI DGX Spark\n\nThis model is a scale up of the Laguna XS 2.1 architecture and vLLM runs it out of the box. Keep your existing Laguna serve setup, and the poolside_v1 tool-call and reasoning parsers already work.","created_at":"2026-07-21T17:18:44.000Z","media":[],"metrics":{"likes":2,"retweets":1,"replies":1,"views":164},"quoted_tweet_id":"2079613777343848465"}]},{"root_id":"2079614410969301331","url":"https://x.com/NVIDIAAI/status/2079614410969301331","author":{"username":"NVIDIAAI","display_name":"NVIDIA AI"},"created_at":"2026-07-21T17:08:07.000Z","lang":"en","categories":["local","models"],"models":["laguna-s-2-1"],"posts":[{"id":"2079614410969301331","text":"Congrats to the @poolsideai team on their latest release, Laguna S 2.1.\n\nIt’s open-weight, delivers way beyond its size, runs great locally and you can customize it with NVIDIA NeMo. \n\nGo try it out on @OpenRouter or download from @huggingface 👉https://t.co/WOIvTVC7wO","created_at":"2026-07-21T17:08:07.000Z","media":[],"metrics":{"likes":29,"retweets":4,"replies":2,"views":2508},"quoted_tweet_id":"2079613777343848465"}]},{"root_id":"2079613777343848465","url":"https://x.com/poolsideai/status/2079613777343848465","author":{"username":"poolsideai","display_name":"Poolside"},"created_at":"2026-07-21T17:05:36.000Z","lang":"en","categories":["agents","context","hardware","models","performance"],"models":["laguna-s-2-1"],"posts":[{"id":"2079613777343848465","text":"Today we're releasing Laguna S 2.1, our most capable model to date.\n\nIt's a 118B total parameter Mixture-of-Experts model with 8B activated per token, a context window of up to 1M tokens, and thinking and no-thinking modes.\n\nCapable enough to hold its own against models many times its size. Small enough to run on a single @NVIDIAAI DGX Spark.\n\nLaguna S 2.1 is fully open under OpenMDW-1.1, with weights available today on @huggingface \n\nhttps://t.co/xxGeAgo35R","created_at":"2026-07-21T17:05:36.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079612009239175168/vid/avc1/1920x1080/jZO91Dm6-PiUmQYT.mp4?tag=29"}],"metrics":{"likes":145,"retweets":40,"replies":21,"views":14874},"quoted_tweet_id":null},{"id":"2079614359446172033","text":"Laguna S 2.1 is, as far as we can measure, the most capable agentic coding model in its weight class.\n\nOn Terminal-Bench 2.1 it scores 70.2, sitting beside models 5–25x its size and ahead of several of them.\n\nAnd on DeepSWE from @datacurve, the hardest long-horizon benchmark we ran, Laguna S 2.1 scores 40.4, outperforming some open models with more than 1T parameters.\n\nFor every score we publish today, we're releasing the full trajectory of every trial in the final evaluation set at https://t.co/2rdFzGic0b","created_at":"2026-07-21T17:07:55.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNxGIMGXAAANvSU.jpg:orig"}],"metrics":{"likes":43,"retweets":5,"replies":3,"views":1409},"quoted_tweet_id":null}]},{"root_id":"2079611556946710818","url":"https://x.com/superalesha/status/2079611556946710818","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-21T16:56:47.000Z","lang":"en","categories":["agents"],"models":[],"posts":[{"id":"2079611556946710818","text":"i just sent one of my codex resets to @elonmusk.\n\nthe desktop app now lets you gift a full reset to another account. if you are not ready to use one, deferral points move its expiry date.\n\n@thsottiaux got the other reset. they both ship more than me. seemed fair. https://t.co/UPq8ATHpVj","created_at":"2026-07-21T16:56:47.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079611346367516672/vid/avc1/1920x1080/0MVeYjXmS7chiEBB.mp4?tag=29"}],"metrics":{"likes":13,"retweets":0,"replies":3,"views":790},"quoted_tweet_id":null},{"id":"2079611559354216692","text":"ok, this is not an openai feature. i patched it into codex desktop for the video.\n\nthe app is electron and the ui lives in an asar bundle, so changing the screen was much easier than it should have been. there is no backend. no reset was really transferred.\n\nstill think openai should steal it. @thsottiaux ship it. everyone else: would you send a spare reset to a friend or hoard it?","created_at":"2026-07-21T16:56:47.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":170},"quoted_tweet_id":null}]},{"root_id":"2079606688122884196","url":"https://x.com/TheZachMueller/status/2079606688122884196","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-21T16:37:26.000Z","lang":"en","categories":["models","quantization","runtime"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079606688122884196","text":"End result using `vllm bench serve`. Not too terrible for a wonky setup. Eventually I'll set everyone up on PCIe5 and can do a proper test (or maybe just rent 6 for a day). But very usable GLM 5.2 NVFP4 at home https://t.co/FkoIzcZIB2","created_at":"2026-07-21T16:37:26.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNw_FMOWwAAJxy5.png:orig"}],"metrics":{"likes":6,"retweets":2,"replies":1,"views":452},"quoted_tweet_id":"2079373432311910445"}]},{"root_id":"2079597687901442404","url":"https://x.com/TheZachMueller/status/2079597687901442404","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-21T16:01:40.000Z","lang":"en","categories":["agents"],"models":[],"posts":[{"id":"2079597687901442404","text":"It’s a good day to have 70% of your daily workflow be Codex https://t.co/rPtxRt0uls","created_at":"2026-07-21T16:01:40.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNw3EfBXUAERVku.jpg:orig"}],"metrics":{"likes":9,"retweets":0,"replies":0,"views":340},"quoted_tweet_id":null}]},{"root_id":"2079596512506438072","url":"https://x.com/aijoey/status/2079596512506438072","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-21T15:57:00.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2079596512506438072","text":"been running NVIDIA Cosmos3 Nano on my DGX Spark\n\ntoday i moved the workflow to Cosmos3 Edge and built a local UI with Hermes agent around it:\n\none road image → 10s progressive rockfall → Locate Anything flagged the debris.\n\nexploring rare, dangerous scenarios without waiting for them to happen.\n\nsimple edge test. more to come.\n\ntip: low step preview before committing to the 50 step render.\n\ncc: @NVIDIAAI @liu_mingyu @NousResearch","created_at":"2026-07-21T15:57:00.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079537461848375296/vid/avc1/1280x960/nWTMZ9egZm4Z2_3U.mp4?tag=29"}],"metrics":{"likes":2,"retweets":1,"replies":0,"views":54},"quoted_tweet_id":null}]},{"root_id":"2079590490400510392","url":"https://x.com/googlegemma/status/2079590490400510392","author":{"username":"googlegemma","display_name":"Google Gemma"},"created_at":"2026-07-21T15:33:04.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079590490400510392","text":"Extremely fast multimodal inference!\n\nDamage Scout uses Gemma 4 on @cerebras running at an impressive 2,300+ toks/s! It analyzes rental car walkaround videos and generates annotated damage reports with box coordinates in <6 seconds. https://t.co/4wI6e17Odp","created_at":"2026-07-21T15:33:04.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079588823126994944/vid/avc1/1080x1080/PERGsztRaiKeKM4I.mp4?tag=29"}],"metrics":{"likes":117,"retweets":4,"replies":3,"views":7241},"quoted_tweet_id":null}]},{"root_id":"2079582939373863353","url":"https://x.com/NVIDIAAI/status/2079582939373863353","author":{"username":"NVIDIAAI","display_name":"NVIDIA AI"},"created_at":"2026-07-21T15:03:04.000Z","lang":"en","categories":["hardware","models"],"models":["deepseek-v3"],"posts":[{"id":"2079582939373863353","text":"NVIDIA Blackwell Ultra just set a world record for pre-training DeepSeek-V3 671B, achieving 1,648 TFLOPs per GPU.\n\nThis is about 3x the earlier delivered performance of the previous generation, and is a result of our extreme co-design and continuous software optimization across popular frameworks including Megatron-Core, TorchTitan, and JAX.","created_at":"2026-07-21T15:03:04.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNwpqS9XsAAL_4p.jpg:orig"}],"metrics":{"likes":382,"retweets":50,"replies":26,"views":21543},"quoted_tweet_id":null}]},{"root_id":"2079573784672211051","url":"https://x.com/danielhanchen/status/2079573784672211051","author":{"username":"danielhanchen","display_name":"Daniel Han"},"created_at":"2026-07-21T14:26:41.000Z","lang":"en","categories":["agents","hardware","performance","quantization"],"models":[],"posts":[{"id":"2079573784672211051","text":"My 2hr workshop on Open vs Closed models, reward hacking, benchmaxxing & RL is out!\n\n1. Closed vs open models\n2. Throughput maxxing but accuracy minimizing\n3. Benchmaxxing & cheating\n4. Distillation & RL\n5. Stopping reward hacking\n6. @UnslothAI Dynamic Quants\n\nDetails:\n1. If reasoning wasn't discovered via o1-preview - would all AI progress grind to a halt via a S shape? Reasoning made doubling times now 3.5 months instead of 7 - so just wait 3.5 months for the next best model! Ways folks might regulate open source - like a driver's license for AI\n\n2. Inference providers maximize throughput and speed, but accuracy degrades - OpenRouter publishes stats on accuracy, and the largest gaps can be 20% or more!\n\n3. METR, WeirdML, Deep-SWE, FrontierCode, SWE Bench Pro - what are good & bad benchmarks? False positive / false negative rates - how about daily benchmarking since Codex & Claude have perf regressions. How can we regressions to predict next model releases as well?\n\n4. Open model labs use distillation partially, but need RL to create the reasoning traces to complete the process - hard vs soft distillation and how to automate RL.\n\n5. Common examples for reward hacking + ways to stop it - internet filtering, classification systems, timing / editing global variables, real world examples + more!\n\n6. Why software matters more than hardware - the limits of FP4 and GPUs vs ASICs and torch.compile vs kernels and megakernels and more, and why quantization and memory optimizations are important\n\nhttps://t.co/tgHm8xpLwp","created_at":"2026-07-21T14:26:41.000Z","media":[],"metrics":{"likes":127,"retweets":26,"replies":10,"views":11420},"quoted_tweet_id":null}]},{"root_id":"2079570757391204587","url":"https://x.com/vanstriendaniel/status/2079570757391204587","author":{"username":"vanstriendaniel","display_name":"Daniel van Strien"},"created_at":"2026-07-21T14:14:39.000Z","lang":"en","categories":["hardware","models"],"models":["qwen3-5-9b"],"posts":[{"id":"2079570757391204587","text":"Making Meta's \"Recycling the Web\" dataset likely cost ~$369k in 2025. Today the same data would cost ~$28k. I checked for $11!\n\nThe data recipe rewrites discarded low-quality web pages from @CommonCrawl into pre-training data with an LLM. I got an agent to re-implement it with datatrove on @huggingface Jobs.\n\nIn 2026, a 9B model beats the paper's own 70B teacher:\n\n→ Qwen3.5-9B clears the quality gate 61.7% of the time vs the 70B's 49.6%\n→ blind paired judges prefer the 9B's rewrites 20-4\n\nWhat it cost:\n\n→ whole PoC: ~$11 in GPU time, 4 models compared\n→ the 9B arm: $0.60 for 600 docs - about $0.64 per million kept tokens\n→ at that rate, Meta's released 44.4B-token set would be ≈ $28k to reproduce. At the 70B teacher's rate ($8.30/M): ≈ $369k (Naive linear extrapolation.)\n\nThe original: filter never released, only part of the data shipped.\n\nThis version: everything's open, including the pipeline, all 1,953 rewrites, cost receipts, and a 30s CPU script that re-scores every rewrite. If you don't trust my agent's eval, run the check yourself!\n\nFeel free to run this at scale: REPRODUCE.md in the repo walks the whole pipeline, and scaling up is two numbers in the executor. If you rebuild the full 44.4B-token set, I'd love to see it! \n\nhttps://t.co/fANTFuGzw5","created_at":"2026-07-21T14:14:39.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNwd3QOXcAAjJEA.jpg:orig"}],"metrics":{"likes":13,"retweets":3,"replies":2,"views":417},"quoted_tweet_id":null}]},{"root_id":"2079547208592982489","url":"https://x.com/StefanMaier/status/2079547208592982489","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-21T12:41:05.000Z","lang":"en","categories":["hardware","models"],"models":["deepseek-v4-flash"],"posts":[{"id":"2079547208592982489","text":"Wanted to have some fun with my @NousResearch Hermes Agent:\nBuilt a Agent-chess game.\n\nI can generate a link out of it, send it to my agent Ares and he knows where to go and what to do.\n\nThe „brain“ I‘m currently challenging is Deepseek v4 Flash on my @NVIDIAAI DGX Spark Cluster 😀\n\nPlaying while chatting with his about it makes some fun 🤓","created_at":"2026-07-21T12:41:05.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNwJKY0XQAAeJi9.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HNwJKYvXgAAHIJf.jpg:orig"}],"metrics":{"likes":1,"retweets":0,"replies":0,"views":26},"quoted_tweet_id":null}]},{"root_id":"2079543927141486805","url":"https://x.com/mr_r0b0t/status/2079543927141486805","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-21T12:28:02.000Z","lang":"en","categories":["models"],"models":["qwen3-6-27b"],"posts":[{"id":"2079543927141486805","text":"Fable is out here throwing refusals and moving to API pricing most people can’t afford.\nMeanwhile, Qwen-3.8 Max is generating 50,000 reasoning traces so we can train Qwen3.6-27B 🔥\nTime to send Anthropic a message!\n#BoycottFable https://t.co/CEAJzyUSR2","created_at":"2026-07-21T12:28:02.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNwGLnnW4AAf_LV.jpg:orig"}],"metrics":{"likes":138,"retweets":14,"replies":9,"views":4507},"quoted_tweet_id":null}]},{"root_id":"2079542489808036080","url":"https://x.com/TheAhmadOsman/status/2079542489808036080","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-21T12:22:20.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2079542489808036080","text":"DROP EVERYTHING\n\n- Why hardware prices, including used RTX 3090s from 2020, are getting more expensive\n\n- Whether frontier intelligence is limited data centers only\n\n- Why you should be bullish on Local / Opensource AI\n\nYou wanna watch this, bookmark it https://t.co/Q3OH95BMzI","created_at":"2026-07-21T12:22:20.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079405880332664832/vid/avc1/1920x1080/FFW_4h45TzKxk8pl.mp4?tag=29"}],"metrics":{"likes":38,"retweets":6,"replies":2,"views":1619},"quoted_tweet_id":null}]},{"root_id":"2079536387859067079","url":"https://x.com/0xSero/status/2079536387859067079","author":{"username":"0xSero","display_name":"0xSero"},"created_at":"2026-07-21T11:58:05.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079536387859067079","text":"Applied this method to Qwen3.6-35B buys you more context on 24gb/32gb\n\nhttps://t.co/LcDa3Ykp3m","created_at":"2026-07-21T11:58:05.000Z","media":[],"metrics":{"likes":173,"retweets":5,"replies":3,"views":16351},"quoted_tweet_id":"2079230064840106173"}]},{"root_id":"2079486068320088553","url":"https://x.com/MiaAI_lab/status/2079486068320088553","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-21T08:38:08.000Z","lang":"en","categories":["concurrency","context","hardware","models","performance","runtime"],"models":[],"posts":[{"id":"2079486068320088553","text":"sparkDash 1.1.7 ⚡\n\n✨ Run decode toks/s benchmarks on the LLM panel — multi-concurrency (default 1–2), real server + per-stream tok/s, TTFT.\n\n✨ vLLM ops tiles (KV cache, queue, TTFT p95, preempts), better GB10 GPU metrics under Docker, and mobile-friendly dialogs.\n\n✨ Lots of bug fixes.\n\n→ https://t.co/0UN5t388W6","created_at":"2026-07-21T08:38:08.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079484858854690816/vid/avc1/2574x1466/nU1kZO1aRPynyhwO.mp4?tag=29"}],"metrics":{"likes":68,"retweets":2,"replies":11,"views":6020},"quoted_tweet_id":null}]},{"root_id":"2079468885544149143","url":"https://x.com/StefanMaier/status/2079468885544149143","author":{"username":"StefanMaier","display_name":"Cesarus"},"created_at":"2026-07-21T07:29:51.000Z","lang":"en","categories":["hardware","local","models"],"models":[],"posts":[{"id":"2079468885544149143","text":"What I’m wondering right now:\n\nWe love the @NVIDIAAI DGX Spark, the new @AMD units, or our fast GPUs for local LLM/AI workloads.\n\nBut where do you draw the line on investment? For some, it’s a hobby (like for me), while others use it for production work (myself included).\n\nHow far are you taking it, and what are your plans?\nWhat are your operating costs (electricity) like?\n\nPersonally, I’m aiming to get two more Sparks. My \"Goliath\" GPU server has likely reached the end of its useful life—if only because of its power consumption...","created_at":"2026-07-21T07:29:51.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNvB7eIXsAAdqwN.jpg:orig"}],"metrics":{"likes":13,"retweets":0,"replies":7,"views":1241},"quoted_tweet_id":null}]},{"root_id":"2079423503124811901","url":"https://x.com/bleysg/status/2079423503124811901","author":{"username":"bleysg","display_name":"Bleys Goodson"},"created_at":"2026-07-21T04:29:31.000Z","lang":"en","categories":["models","performance"],"models":["deepseek-r1"],"posts":[{"id":"2079423503124811901","text":"Amazingly, when I crunched what it takes to serve Kimi K3, I found not only Huawei's next-gen Ascend 950DT on track to beat the Nvidia B300 on the all-important tokens-per-watt metric, but also the current CloudMatrix 384 generation already at parity at the 50 tok/s speeds API providers aim to maintain.\n\nAscend 950DT based on announced specs (with an assumed power figure). Projections calibrated on Huawei's published DeepSeek-R1 serving results; K3 geometry inferred until the 7/27 weight drop.","created_at":"2026-07-21T04:29:31.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNuUFPibwAAoxgI.jpg:orig"}],"metrics":{"likes":22,"retweets":2,"replies":5,"views":6776},"quoted_tweet_id":"2079301851867136422"}]},{"root_id":"2079406563370951064","url":"https://x.com/TheAhmadOsman/status/2079406563370951064","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-21T03:22:12.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079406563370951064","text":"The Desktop Frontier\n\nWe're gonna get Kimi K3 equivalent intelligence running on a single RTX PRO 6000 in less than 18 months\n\nHow? Watch this video if you wanna learn how we get there\n\nBookmark for the future https://t.co/dviz2vONd6","created_at":"2026-07-21T03:22:12.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079405880332664832/vid/avc1/1920x1080/FFW_4h45TzKxk8pl.mp4?tag=29"}],"metrics":{"likes":349,"retweets":29,"replies":23,"views":20030},"quoted_tweet_id":null}]},{"root_id":"2079395288335503636","url":"https://x.com/bleysg/status/2079395288335503636","author":{"username":"bleysg","display_name":"Bleys Goodson"},"created_at":"2026-07-21T02:37:24.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079395288335503636","text":"⚡Save this and read it over a coffee break if you want to understand what everyone else is missing in the market after the Kimi K3 release and Xi Jinping's grand new strategic open weights directive this week.","created_at":"2026-07-21T02:37:24.000Z","media":[],"metrics":{"likes":2,"retweets":0,"replies":0,"views":569},"quoted_tweet_id":"2079387300698669379"}]},{"root_id":"2079387300698669379","url":"https://x.com/bleysg/status/2079387300698669379","author":{"username":"bleysg","display_name":"Bleys Goodson"},"created_at":"2026-07-21T02:05:40.000Z","lang":"en","categories":["agents","context","models"],"models":[],"posts":[{"id":"2079387300698669379","text":"Thanks for these thoughts Nathan. We largely agree: the trailing equilibrium is real and a net good, near-term risks are overhyped, bans make things less safe, and independent evals are the institution worth building. But I strongly reject that open weights as we're seeing them are decelerationist even for the leading labs. Ball is not right here, and even splitting the difference is unfair to the reality. Dean quickly shrunk the claim to open weights decelerating capex on the margin, calling that much straightforwardly true. Even the shrunken version has the sign wrong.\n\nLet me lay it out. Opus 4.6 at its peak commanded 3/4 of Anthropic's API revenue, near 80% with Fast Mode. Beneath it in the same SemiAnalysis mix, the prior Opus generation collapsed from 10% of revenue to 0.2% in five quarters, its price from $35 to $3 per million tokens. Open weights didn't do that. The leader's next release did. The cadence liquidates last cycle's rents before any open model arrives to claim them. Mythos/Fable then more than doubled Anthropic's run rate overnight, and something like 90% of model revenue now sits with three models. Open weights compete for the fumes of the release cadence, capability already repriced down 90% and moved past. The subscriber annuity buys the product, not the weights.\n\nEven if we concede an equilibrium somewhere in between, say open weights a cycle behind account for a full 20% of revenue, notice why the number is bounded: the floor can only tax capability it already has, the frontier delta is priced against displaced wages, and the subscription layer is product economics. Seeing 20% as decelerative fails to grasp what acceleration even is: trading the linear for the exponential. A bounded share is a linear cost, and a linear cost vanishes in the limit.\n\nMeanwhile what that cost buys sits inside the exponent, per your own reporting. Gated DeltaNet: late-2024 academia to frontier scale by mid-2026. KDA: Kimi Linear paper into Olmo Hybrid, Qwen, and Nemotron in a fraction of that. And the channel the buffer frame doesn't price: labs RL on each other's releases. The distillation fights prove frontier outputs are already the industry's teacher signal, harvested through API gates. Weights in the open remove the gate: every lab gets an unmetered teacher, grader, and trace generator. That is not architecture diffusion on an eighteen-month clock. It is post-training convergence on a quarterly one. Each release becomes next cycle's curriculum for everyone, the floor converges faster, and willingness-to-pay concentrates in the one thing the floor can't teach: the current delta.\n\nThat leaves the fundraising channel, where cheaper open models are meant to compress margins, dent valuations, and slow the next raise. It fails empirically for the same reason. Twice now a frontier-class open release has hit the tape, chip stocks have convulsed, and the market has bounced inside forty-eight hours. $1.4T committed against $25B of revenue never priced margin annuities; it prices cadence, position, and a market that doubles as capability climbs. And the tax is levied by labs that raised orders of magnitude less capital and train because they lack inference demand. The deceleration story needs followers who depend on the leader's margins. Your reporting shows they don't. The exponential does not notice the toll.\n\nSo I'd push one step past your closing frame: you call the open-closed dance collective action for the community to steward. It doesn't need stewarding. Every party in it is being paid. The catch-up tier ships because shipping is its only relevance. The leader banks the architectures, the free teachers, and the trained-up cohort.\n\nThe gap you call society's safety buffer doubles as the leader's paid moat. Open models are the slower-starting bigger exponential, agreed; my claim is the leaders ride the same curve. Last-cycle weights in the open accelerate everyone's next release, all compute buildout, and the expansion into higher-value work the leaders capture on their release. Raising the floor makes next quarter's burst bigger by more than it costs last quarter's tail.\n\nStep back and the whole disagreement has the shape of a single inversion. The deceleration story, in every channel, casts the open tier as a cost the leading labs bear. It is the opposite. The open tier is their R&D pipeline, teacher network, customer curriculum, and demand flywheel, financed by their competitors, consuming only rents the cadence already burned. Ball's claim isn't overstated. It's pointed backwards. Open weights aren't the brake. They're nearly-free nitrous. The floor isn't eating the frontier's lunch. It's setting the frontier's table.","created_at":"2026-07-21T02:05:40.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNtum8MaUAAmja2.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HNtvIHvbUAAmPXw.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HNtvsgqbEAEZ973.png:orig"}],"metrics":{"likes":3,"retweets":0,"replies":2,"views":2885},"quoted_tweet_id":"2079232178396033394"},{"id":"2079404488402841642","text":"We can already see what the leaders will compound from the trailing tier: look at what K3 actually is. \"Trailing\" is an average, not a property.\n\nMoonshot sits months behind the frontier overall, and simultaneously first in legal agentic work on Harvey's LAB-AA, in agentic SaaS workflows, in long-context reasoning, in agentic tool use on τ³-Banking, and on both frontend arenas. The floor is jagged, and several of its peaks are currently the ceiling.\n\nOn July 27 every one of those peaks becomes common property. Recall the teacher channel: the leaders don't just get a parity-grade teacher, they get a specialist sitting above them in each of those slices. An RL grader for legal agentic reliability. A trace generator for frontend taste. A long-context curriculum nobody else can produce. Alpha the leaders would otherwise spend a post-training cycle discovering, handed over under a permissive license. Free wins, compounded into the very next release.\n\nThis is what the gains from the common floor are concretely. The concession pays in categories; the cadence collects them. It ran the other way first: K2.6 was #18 a year ago, and K3 climbed to #1 partly on everyone else's exhaust. The compounding runs in both directions, which is the entire engine. Deceleration pictures the floor dragging the frontier down. The leaderboards show the floor raising specific ceilings, and then handing over the ladder.","created_at":"2026-07-21T03:13:58.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNuGeg5awAAFBW4.jpg:orig"},{"type":"photo","url":"https://pbs.twimg.com/media/HNuHSz8bMAAo7n4.jpg:orig"}],"metrics":{"likes":0,"retweets":0,"replies":0,"views":181},"quoted_tweet_id":null}]},{"root_id":"2079373432311910445","url":"https://x.com/TheZachMueller/status/2079373432311910445","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-21T01:10:33.000Z","lang":"en","categories":["agents","models","performance","quantization"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079373432311910445","text":"Where my cursed rig setup has ended after 6.4M Sol tokens over 10 hrs 47 min optimizing GLM 5.2 NVFP4. \n\nCursed topology of DDR4 4x Max-Q + DDR5 2x 6000 Pro, ending with Asymmetric TP2/TP4 + async PP to get ~47 tok/s/user\n\nUnrelated: anyone got one of those codex resets handy? https://t.co/UPyNLVFWWD","created_at":"2026-07-21T01:10:33.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNtrEn6XwAAEMxX.jpg:orig"}],"metrics":{"likes":19,"retweets":1,"replies":2,"views":1047},"quoted_tweet_id":null}]},{"root_id":"2079334539692814514","url":"https://x.com/TheAhmadOsman/status/2079334539692814514","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-20T22:36:01.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079334539692814514","text":"Calling Opensource AI a cybersecurity threat is like saying we should've cancelled the internet because viruses spread through it\n\nClosed models refuse to fix security issues in my code while Kimi K3 helps\n\nWhich one is actually unsafe","created_at":"2026-07-20T22:36:01.000Z","media":[],"metrics":{"likes":384,"retweets":43,"replies":28,"views":7008},"quoted_tweet_id":null}]},{"root_id":"2079324234883407883","url":"https://x.com/BosonJoe/status/2079324234883407883","author":{"username":"BosonJoe","display_name":"Joe Muller"},"created_at":"2026-07-20T21:55:04.000Z","lang":"en","categories":["context","hardware","models","performance","quantization"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079324234883407883","text":"🎆New decode record 🎆\n\nNew: 22.8 tok/s 🔥\nOld: 19.8 tok/s\n\nGLM 5.2 on 2 DGX Sparks\n\nI quantized the dspark speculator to W4A16 (7.1GB -> 4.6GB w/ near identical acceptance)\n\nA smaller drafter + bumping KV cache = 96k context\n\n⚡️⚡️⚡️⚡️⚡️ https://t.co/wn1vwhzPSF","created_at":"2026-07-20T21:55:04.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079317030272958465/vid/avc1/2272x1760/PKn7BEpgjN84aFoo.mp4?tag=29"}],"metrics":{"likes":97,"retweets":10,"replies":13,"views":9222},"quoted_tweet_id":"2077819511730802846"}]},{"root_id":"2079298811726569942","url":"https://x.com/sudoingX/status/2079298811726569942","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-20T20:14:02.000Z","lang":"en","categories":["hardware","models","performance","quantization"],"models":[],"posts":[{"id":"2079298811726569942","text":"the chip ban was supposed to keep china a decade behind. this is the receipt. huawei's atlas 950 superpod, one eflops fp8, 256tb of unified memory, 1024 accelerators, built entirely on ascend silicon because you told them they couldn't buy nvidia.\n\nyou didn't slow them down. you handed them a reason to build their own stack, and they did it in eighteen months. there was no ascend supercluster ecosystem before the export controls forced one into existence.\n\nand the part nobody in dc thought through. you can restrict chips. you cannot restrict open weights. kimi k3 is already trading blows with fable class frontier models on some benchmarks, and it ships open. the model is on every drive that wanted it. what's the tariff on a file?\n\nso the ban gave china a domestic hardware industry and left the models ungovernable. cheap ascend accelerators are going to flood this market, and every builder with a garage gets frontier adjacent ai on hardware they own.\n\nworst moat i've ever watched get built. it funded the competitor and freed the software. i'm not even mad, cheap hardware is the only ideology i've got.","created_at":"2026-07-20T20:14:02.000Z","media":[],"metrics":{"likes":288,"retweets":35,"replies":12,"views":17973},"quoted_tweet_id":"2078028640059859312"}]},{"root_id":"2079270345031192605","url":"https://x.com/TheAhmadOsman/status/2079270345031192605","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-20T18:20:55.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079270345031192605","text":"Local AI was never this EASY\n\n> Install ODS\n> Let it detect your hardware\n> It will download the best model for your hardware\n> And then start local inference and Open WebUI for you\n\nWith ODS, you can\n\n> Add voice, agents like Hermes, workflows, RAG, search, image generation, and more\n> Manage the whole stack from one dashboard\n\nNow your PC, Mac, or Linux box is a private AI server\n\nNo cloud required\nNo subscription required\nYour prompts and data stay on your machine unless you choose otherwise\n\nWe're gonna make Local AI The Default","created_at":"2026-07-20T18:20:55.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNsNW_kXoAEA7NP.jpg:orig"}],"metrics":{"likes":143,"retweets":18,"replies":13,"views":10274},"quoted_tweet_id":null}]},{"root_id":"2079266316267946445","url":"https://x.com/BosonJoe/status/2079266316267946445","author":{"username":"BosonJoe","display_name":"Joe Muller"},"created_at":"2026-07-20T18:04:55.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079266316267946445","text":"I predict that soon local AI rigs will be treated the same as cars\n\nHaving one will be normal\n\nYou'll pay it off over 3-5 years\n\nRig envy will be everywhere\n\nAnd you'll \"take it to work\"","created_at":"2026-07-20T18:04:55.000Z","media":[],"metrics":{"likes":51,"retweets":4,"replies":4,"views":4045},"quoted_tweet_id":"2077123292352204943"}]},{"root_id":"2079265527042830345","url":"https://x.com/superalesha/status/2079265527042830345","author":{"username":"superalesha","display_name":"Alexey Fateev"},"created_at":"2026-07-20T18:01:47.000Z","lang":"en","categories":["models","performance","quantization","runtime"],"models":[],"posts":[{"id":"2079265527042830345","text":"so, I rented the RTX PRO 6000 again. another 10 hours.\n\nvLLM vs SGLang, for real this time. same nvidia qwen 3.6 checkpoint (27b, nvfp4), fp8 kv, byte-for-byte identical prompts, 87 valid measured points. no more excuses.\n\nthroughput on a cold start, 8k prompt, 32 users: vllm - 96.60 tokens/s. sglang - 96.59.\n\na tie... accurate to the second decimal place. lol.\n\nthe real fun starts when you look at whats happening around it","created_at":"2026-07-20T18:01:47.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNmxz_rWMAAk1uF.jpg:orig"}],"metrics":{"likes":36,"retweets":2,"replies":5,"views":3324},"quoted_tweet_id":null}]},{"root_id":"2079240759463604505","url":"https://x.com/MiaAI_lab/status/2079240759463604505","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-20T16:23:22.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079240759463604505","text":"Kimi K3 vs Qwen 3.8\n\nI asked both to create an interactive solar system. In this specific test I prefer Qwen 3.8 quite a bit more.\n\nThe zoom in/out is smoother, and the orbiting looks and feels a lot better.\n\nI’ve also noticed that Kimi K3 always wants to check itself and usually runs a headless browser to find issues before it’s done. Qwen 3.8 doesn’t.\n\nLive demos here:\nhttps://t.co/Ggz4CC851r\n\nFull prompt below.","created_at":"2026-07-20T16:23:22.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079237462124052480/vid/avc1/3840x2160/oS2Xvo3io6PbC7hA.mp4?tag=29"}],"metrics":{"likes":257,"retweets":16,"replies":30,"views":24797},"quoted_tweet_id":null}]},{"root_id":"2079236204743053592","url":"https://x.com/NVIDIAAI/status/2079236204743053592","author":{"username":"NVIDIAAI","display_name":"NVIDIA AI"},"created_at":"2026-07-20T16:05:16.000Z","lang":"en","categories":["agents","hardware","local","models"],"models":[],"posts":[{"id":"2079236204743053592","text":"Introducing Cosmos 3 Edge: our open frontier world model built to run on-device.\n\nCosmos 3 Edge helps robots learn and act, autonomous vehicles understand road scenes and predict intent, and vision AI agents reason across live video for smart infrastructure.\n\nWith 4B parameters and a 2B Nemotron-based reasoner, you can run it on DGX Spark, NVIDIA Jetson, and more.","created_at":"2026-07-20T16:05:16.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079236168617467905/vid/avc1/1280x720/gG9jYIHEK_kL5oW2.mp4?tag=14"}],"metrics":{"likes":1578,"retweets":171,"replies":44,"views":125078},"quoted_tweet_id":null}]},{"root_id":"2079234161626447885","url":"https://x.com/LottoLabs/status/2079234161626447885","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-20T15:57:09.000Z","lang":"en","categories":["hardware","models"],"models":[],"posts":[{"id":"2079234161626447885","text":"3090 + tailscale + qwen 27b 3.8 + pi + tmux\n\nAnd you’re free https://t.co/Vf2vs7GbwO","created_at":"2026-07-20T15:57:09.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNrscnYaoAAWL60.jpg:orig"}],"metrics":{"likes":9,"retweets":2,"replies":0,"views":124},"quoted_tweet_id":null}]},{"root_id":"2079233513430438010","url":"https://x.com/LottoLabs/status/2079233513430438010","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-20T15:54:34.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079233513430438010","text":"Local ai is neat because you don’t need a phd to tinker around","created_at":"2026-07-20T15:54:34.000Z","media":[],"metrics":{"likes":16,"retweets":0,"replies":2,"views":426},"quoted_tweet_id":null}]},{"root_id":"2079222227958899018","url":"https://x.com/RedHat_AI/status/2079222227958899018","author":{"username":"RedHat_AI","display_name":"Red Hat AI"},"created_at":"2026-07-20T15:09:43.000Z","lang":"en","categories":["agents","context","models"],"models":["gemma-4-31b-it","gemma-4-31b"],"posts":[{"id":"2079222227958899018","text":"Two powerful new models now available on Red Hat AI Inference on IBM Cloud. Two different strengths.\n\n☑️ Granite 4.1 30B for enterprise assistants, RAG, code generation, and agentic workflows\n\n☑️ Gemma 4 31B IT for multimodal reasoning, image understanding, and long-context applications\n\nBoth are available via familiar OpenAI-compatible APIs.\n\nNice work by the team expanding the model catalog and giving developers more choice without changing how they build and integrate AI applications.\n\nGetting started with these models on Red Hat AI Inference on IBM cloud: https://t.co/ZS252DkC0C","created_at":"2026-07-20T15:09:43.000Z","media":[],"metrics":{"likes":3,"retweets":1,"replies":0,"views":234},"quoted_tweet_id":null}]},{"root_id":"2079221538256007629","url":"https://x.com/MiaAI_lab/status/2079221538256007629","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-20T15:06:59.000Z","lang":"en","categories":["hardware","models"],"models":[],"posts":[{"id":"2079221538256007629","text":"sparkDash ver 1.0.5 is out ⚡️\n\n- Multiple LLM ports — monitor several LLM servers on different ports simultaneously\n- GPU processes — see top GPU processes by VRAM usage with name and memory\n- System uptime inline on each Spark header\n\nLots of stuff planned to make this the ultimate dashboard for your DGX Sparks!\n\nhttps://t.co/8H3h5wjPOh","created_at":"2026-07-20T15:06:59.000Z","media":[],"metrics":{"likes":7,"retweets":0,"replies":1,"views":1073},"quoted_tweet_id":"2078944033838116876"}]},{"root_id":"2079216310655803449","url":"https://x.com/MiaAI_lab/status/2079216310655803449","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-20T14:46:13.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079216310655803449","text":"Just did a quick test with Qwen 3.8 Preview — it seems to think for 10 minutes straight and still no results. \n\nWasn’t even a hard prompt either… 🤷♀️ https://t.co/4F0vX6yK7q","created_at":"2026-07-20T14:46:13.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079215362596384769/vid/avc1/2302x1540/HZJCGWsv3GnZi3Lv.mp4?tag=29"}],"metrics":{"likes":7,"retweets":0,"replies":2,"views":135},"quoted_tweet_id":null}]},{"root_id":"2079211881185619982","url":"https://x.com/vanstriendaniel/status/2079211881185619982","author":{"username":"vanstriendaniel","display_name":"Daniel van Strien"},"created_at":"2026-07-20T14:28:37.000Z","lang":"en","categories":["agents","models"],"models":[],"posts":[{"id":"2079211881185619982","text":"Your coding agent can now do semantic search to find models and datasets on the @huggingface Hub!\n\nmeans you can hand it a prompt like:\n\n\"Find recent datasets of video game gameplay useful for training game-playing agents. Curate the best ones into a collection with a brief description for each.\"\n\n...and it goes and does it.\n\nI built hf find (a hf extension) to give agents a way to look up what's actually on the Hub: semantic search over 1.17M model and dataset cards, ranked by meaning. \n\nIt detects when an agent is driving and switches to TSV with hints on stderr (--format json for scripts), and every result ends with the next command to run, so it can go from search to find-similar to the card on its own. The index refreshes daily from a scheduled HF Job for about 5 cents a day.\n\nStill some rough edges but quite cool to see agents use tools like this! \n\nhf extensions install davanstrien/hf-find","created_at":"2026-07-20T14:28:37.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2079210299433869312/vid/avc1/1154x646/3xYhpgpu96tonSJN.mp4?tag=29"}],"metrics":{"likes":0,"retweets":0,"replies":1,"views":13},"quoted_tweet_id":null}]},{"root_id":"2079207457788952944","url":"https://x.com/UnslothAI/status/2079207457788952944","author":{"username":"UnslothAI","display_name":"Unsloth AI"},"created_at":"2026-07-20T14:11:02.000Z","lang":"en","categories":["agents","hardware","local","models","quantization"],"models":[],"posts":[{"id":"2079207457788952944","text":"Introducing Unsloth for AMD 🚀\nYou can now train & run LLMs on your AMD hardware\n\n• We collaborated with AMD to enable you to train & run 500+ models on AMD GPUs\n• Works on Windows, WSL, Linux\n• Train Qwen, Gemma on 3GB VRAM\n\nGitHub: https://t.co/2kXqhhvLsb\n\nWorks on Radeon, Instinct, Ryzen and data center GPUs with up to 2× faster with 70% less VRAM and no accuracy loss via our custom Triton kernels and math algorithms. We also support optimized ROCm builds for GGUF & Safetensors inference.\n\nUnsloth is an open-source local UI for faster LLM training and inference, with tool-call healing, code execution, secure web search, remote APIs, and HTTPS deployment. Connect local models to Claude Code, Codex agents and run the latest Kimi, GLM, DeepSeek, Qwen3.6, and Gemma 4 models.\n\n🔗Blog + Guide: https://t.co/U9LqyRjFdj","created_at":"2026-07-20T14:11:02.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNrRih3bMAAoJus.jpg:orig"}],"metrics":{"likes":54,"retweets":11,"replies":4,"views":1271},"quoted_tweet_id":null}]},{"root_id":"2079197700726071563","url":"https://x.com/MiaAI_lab/status/2079197700726071563","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-20T13:32:16.000Z","lang":"it","categories":["hardware"],"models":[],"posts":[{"id":"2079197700726071563","text":"Insane 8x DGX Spark cluster setup 🔥","created_at":"2026-07-20T13:32:16.000Z","media":[],"metrics":{"likes":12,"retweets":2,"replies":1,"views":1230},"quoted_tweet_id":"2079158726658060652"}]},{"root_id":"2079181535626842293","url":"https://x.com/mr_r0b0t/status/2079181535626842293","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-20T12:28:02.000Z","lang":"en","categories":["context"],"models":[],"posts":[{"id":"2079181535626842293","text":"TIL the GPT5.6-XXX-Pro models in Nous Portal have a 1M context window 👀 https://t.co/nlgqrDfTUv","created_at":"2026-07-20T12:28:02.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNq8g6RW4AEha8Y.png:orig"}],"metrics":{"likes":25,"retweets":1,"replies":4,"views":1503},"quoted_tweet_id":null}]},{"root_id":"2079180164399878288","url":"https://x.com/TheAhmadOsman/status/2079180164399878288","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-20T12:22:35.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079180164399878288","text":"@TheZachMueller Yeah, I agree. K2.7 and GLM 5.2 came out in the same time so I consider one of them representative of the other.\n\nVery curious about the data you are talking about though, is it public?","created_at":"2026-07-20T12:22:35.000Z","media":[],"metrics":{"likes":6,"retweets":0,"replies":1,"views":265},"quoted_tweet_id":null}]},{"root_id":"2079173030312354224","url":"https://x.com/TheAhmadOsman/status/2079173030312354224","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-20T11:54:14.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079173030312354224","text":"I consider all code written before GPT 5.6 / GLM 5.2 / Kimi K3 as heavy tech debt\n\nAfter those releases we're closer to normal tech debt\n\nPretty much most of the code produced between 2024-2027 will be refactored in full by Agents (my guess 2028-2029, memory arch improvements)","created_at":"2026-07-20T11:54:14.000Z","media":[],"metrics":{"likes":48,"retweets":2,"replies":15,"views":1847},"quoted_tweet_id":null}]},{"root_id":"2079161991952797707","url":"https://x.com/antirez/status/2079161991952797707","author":{"username":"antirez","display_name":"antirez"},"created_at":"2026-07-20T11:10:22.000Z","lang":"en","categories":["concurrency","models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2079161991952797707","text":"I just pushed a new version of DwarfStar that is likely to break stuff, too many new things: 1. micro-batching. 2. CUDA multi device support. 3. GLM5.2 4. Metal tensor parallel execution (via RDMA). 4. DSpark speculative execution. 5. Many other things probably.","created_at":"2026-07-20T11:10:22.000Z","media":[],"metrics":{"likes":90,"retweets":6,"replies":7,"views":6290},"quoted_tweet_id":null}]},{"root_id":"2079156452719989000","url":"https://x.com/antirez/status/2079156452719989000","author":{"username":"antirez","display_name":"antirez"},"created_at":"2026-07-20T10:48:21.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2079156452719989000","text":"It is always a good moment to remember that the recent results disprove another conjecture as well: the one of the many AI scientists that were LLM skeptics. People make errors, but remember that AI is not fully understood, and experts of the field make a lot of blatant errors.","created_at":"2026-07-20T10:48:21.000Z","media":[],"metrics":{"likes":156,"retweets":9,"replies":11,"views":8520},"quoted_tweet_id":null}]},{"root_id":"2079137278840955081","url":"https://x.com/TheAhmadOsman/status/2079137278840955081","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-20T09:32:10.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2079137278840955081","text":"ODS will \"save hundreds of hours\" in setup and infra time for you when you decide on migrating to Local AI btw https://t.co/IUronMuxy7","created_at":"2026-07-20T09:32:10.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNogQE_W8AAnzIB.jpg:orig"}],"metrics":{"likes":21,"retweets":1,"replies":5,"views":1098},"quoted_tweet_id":"2079008486818725912"}]},{"root_id":"2079134549531935160","url":"https://x.com/TheZachMueller/status/2079134549531935160","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-20T09:21:19.000Z","lang":"en","categories":["agents","performance"],"models":[],"posts":[{"id":"2079134549531935160","text":"Update: codex ended the night at doubling tok/s/user\n\nWhat","created_at":"2026-07-20T09:21:19.000Z","media":[],"metrics":{"likes":5,"retweets":0,"replies":2,"views":458},"quoted_tweet_id":"2079033288463671719"}]},{"root_id":"2079033288463671719","url":"https://x.com/TheZachMueller/status/2079033288463671719","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-20T02:38:57.000Z","lang":"en","categories":["hardware","performance"],"models":[],"posts":[{"id":"2079033288463671719","text":"Update: ending the night at 20 tok/s/user (10GbE, 4/2 split GPUs between nodes). \n\nThis is pre-optimizing + using spec decode.","created_at":"2026-07-20T02:38:57.000Z","media":[],"metrics":{"likes":16,"retweets":1,"replies":1,"views":8817},"quoted_tweet_id":"2078975786812420373"}]},{"root_id":"2079005762144321901","url":"https://x.com/alexocheema/status/2079005762144321901","author":{"username":"alexocheema","display_name":"Alex Cheema"},"created_at":"2026-07-20T00:49:34.000Z","lang":"en","categories":["hardware","local","models","performance","quantization"],"models":[],"posts":[{"id":"2079005762144321901","text":"I'll be sharing a first look at local dot ai with @huggingface on Tuesday.\n\nWe benchmarked every model, quantization, harness, MTP setting, EVERYTHING, on every local device - and we're opening it up for free on local dot ai.\n\nEvery single data point on local dot ai has 700+ tasks behind it. We measure what an end user will actually experience - intelligence and end-to-end time to complete tasks - to answer questions like:\n- What device is the best value for Local AI?\n- What model should I run on my MacBook or DGX Spark?\n- How do I cluster devices to boost performance / run larger models?\n\nSo many insights to cover. Be there Tuesday.","created_at":"2026-07-20T00:49:34.000Z","media":[],"metrics":{"likes":120,"retweets":15,"replies":19,"views":14952},"quoted_tweet_id":"2078878406335201738"}]},{"root_id":"2078975786812420373","url":"https://x.com/TheZachMueller/status/2078975786812420373","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-19T22:50:27.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2078975786812420373","text":"Electric bill is about to 2x. Let's try and run GLM 5.2 at home. https://t.co/CNrEZtVamE","created_at":"2026-07-19T22:50:27.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNoBcH7WYAAqDeS.png:orig"}],"metrics":{"likes":16,"retweets":0,"replies":4,"views":2133},"quoted_tweet_id":null}]},{"root_id":"2078955932445536405","url":"https://x.com/aijoey/status/2078955932445536405","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-19T21:31:34.000Z","lang":"en","categories":["concurrency"],"models":[],"posts":[{"id":"2078955932445536405","text":"For everyone just getting on herdr. Wait till you use via termius/tailscale\n\nI've been using it for 2 weeks now. Super impress and it's made it much easier to maneuver. \n\nDetach and come back:\nPress prefix+q or simply close your terminal window. The Herdr server and every agent keep running. Run herdr again to reattach to the same session.\n\nTo actually end the session and stop its panes:\n\nherdr server stop","created_at":"2026-07-19T21:31:34.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2078955902762180608/vid/avc1/1320x2868/XRrncwp2XnRRKePg.mp4?tag=29"}],"metrics":{"likes":50,"retweets":4,"replies":6,"views":4406},"quoted_tweet_id":null}]},{"root_id":"2078953366659518970","url":"https://x.com/Ex0byt/status/2078953366659518970","author":{"username":"Ex0byt","display_name":"Eric"},"created_at":"2026-07-19T21:21:22.000Z","lang":"en","categories":["context","hardware","models","performance","quantization","runtime"],"models":["deepseek-v4-flash"],"posts":[{"id":"2078953366659518970","text":"@KingBreton78876 we were at 13 tok/s a few weeks ago if you recall - patience my friend, soon. We'll release a one click set up image once I clean up all of the previously closed experiments.","created_at":"2026-07-19T21:21:22.000Z","media":[],"metrics":{"likes":20,"retweets":0,"replies":2,"views":1327},"quoted_tweet_id":null},{"id":"2078955117638734316","text":"@ic4y00 one of the best models sir.","created_at":"2026-07-19T21:28:19.000Z","media":[],"metrics":{"likes":4,"retweets":0,"replies":0,"views":853},"quoted_tweet_id":null},{"id":"2078999113920663665","text":"@CliffordManass5 the kernels should be transferable, cant say for sure, don't have the strix halo in the lab.","created_at":"2026-07-20T00:23:09.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":638},"quoted_tweet_id":null},{"id":"2079237259983770077","text":"Not magic, just well reasoned engineering. \n\nRecipe: DeepSeek-V4-Flash: 48-49 tok/s TP 4 chats on a single GB10 workstation\n\nHardware and model\n\nHardware: 1x GB10 (128 GB unified memory, ~273 GB/s bandwidth)\nModel: DeepSeek-V4-Flash, 284B total params, 13B active, 256 routed experts plus 1 shared per layer, 6 active per token\n\nSGLANG: Serving (pre-packing)\n\nSeparate dense layers from expert layers\nCompress experts with a codebook (shared dictionary plus short indices, ~2 bits/weight; not quantization!)\nCustom CUDA kernel computes directly on compressed weights via dictionary lookup, ~90% of memory bandwidth\n\nMemory\n\nFits fully resident: ~69 GB experts, ~6.5 GB dense, small KV cache, 85-95 GB total\nSSD streaming optional! Use it for faster startup and context headroom; required only for 1.6T V4-Pro, not Flash\n\nSpec-dec (DSPARK)\n\nSmall draft model guesses, big model verifies in one pass, ~3-6 tokens per verification, lossless output!\nCritical: draft input must exactly match its training format (native block size, its own internal states); any mismatch gives zero accepted guesses\n\nBoot (both failures look like OOM but are not)\n\nPre-packed expert file must be mounted at startup or the server dies silently\nWrite a durable ready record at boot or an internal cache reports unhealthy and shuts down\n\nBenchmark\n\ntemperature 1.0, top-p 1.0; (greedy settings inflate acceptance and fake the numbers)\n\nResult:\n\n48-49+ tok/s with spec-dec, up from ~31 baseline from a few weeks ago.","created_at":"2026-07-20T16:09:27.000Z","media":[],"metrics":{"likes":5,"retweets":0,"replies":2,"views":370},"quoted_tweet_id":null},{"id":"2079253759725580378","text":"@JoesInvestments https://t.co/EoV0e9BpwB","created_at":"2026-07-20T17:15:01.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":162},"quoted_tweet_id":"2079237259983770077"},{"id":"2079291721146388603","text":"@with_gene2626 @NeoAIForecast 65k","created_at":"2026-07-20T19:45:52.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":1,"views":43},"quoted_tweet_id":null}]},{"root_id":"2078944033838116876","url":"https://x.com/MiaAI_lab/status/2078944033838116876","author":{"username":"MiaAI_lab","display_name":"Mia"},"created_at":"2026-07-19T20:44:17.000Z","lang":"en","categories":["hardware","local","models","performance","runtime"],"models":["deepseek-v4-flash"],"posts":[{"id":"2078944033838116876","text":"I'm open-sourcing sparkDash ⚡️\n\n- Real-time monitoring for your DGX Sparks \n- Add your sparks from the UI\n- vLLM / llama.cpp / SGLang real-time tok/s\n- Mobile-friendly\n\nBuilt 100% with DeepSeek V4 Flash running locally.\n\nLet me know what you think 🫶\n\n→ https://t.co/8H3h5wjPOh","created_at":"2026-07-19T20:44:17.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNng4ZGWEAAlWSq.jpg:orig"}],"metrics":{"likes":279,"retweets":24,"replies":35,"views":16631},"quoted_tweet_id":null}]},{"root_id":"2078939915824386513","url":"https://x.com/TheAhmadOsman/status/2078939915824386513","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-19T20:27:55.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2078939915824386513","text":"In all seriousness though, Mike is an amazing human being who genuinely just wants to help people by sharing his knowledge & experiences\n\nThis guy led $65m in operations annually and then decided to go all in on Local AI because he saw something\n\nFollow him if you don’t, top guy","created_at":"2026-07-19T20:27:55.000Z","media":[],"metrics":{"likes":359,"retweets":22,"replies":7,"views":51725},"quoted_tweet_id":"2045151957350764794"}]},{"root_id":"2078923846766256316","url":"https://x.com/BosonJoe/status/2078923846766256316","author":{"username":"BosonJoe","display_name":"Joe Muller"},"created_at":"2026-07-19T19:24:04.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2078923846766256316","text":"This is actually amazing and I use it multiple times a day 📊\n\nIf you own an @NVIDIAAI DGX Spark, check it out! https://t.co/iJbfcC9HfO","created_at":"2026-07-19T19:24:04.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2078923571561189376/vid/avc1/1608x1654/-05NwTtrq2Ya0AdQ.mp4?tag=29"}],"metrics":{"likes":195,"retweets":16,"replies":12,"views":16739},"quoted_tweet_id":null}]},{"root_id":"2078917396199280724","url":"https://x.com/Ex0byt/status/2078917396199280724","author":{"username":"Ex0byt","display_name":"Eric"},"created_at":"2026-07-19T18:58:26.000Z","lang":"en","categories":["hardware","models","performance"],"models":["deepseek-v4-flash"],"posts":[{"id":"2078917396199280724","text":"(Fable stands corrected, in shame) Sol + Kimi for the win!\n\nDeepSeek-V4-Flash: ~50 tok/s on 1 DGX Spark, full quality, no pruning. https://t.co/hkayiVUZQg","created_at":"2026-07-19T18:58:26.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNnKdrdWIAACQjd.jpg:orig"}],"metrics":{"likes":104,"retweets":6,"replies":7,"views":8532},"quoted_tweet_id":null}]},{"root_id":"2078908682666950843","url":"https://x.com/mr_r0b0t/status/2078908682666950843","author":{"username":"mr_r0b0t","display_name":"mr-r0b0t"},"created_at":"2026-07-19T18:23:48.000Z","lang":"en","categories":["models"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2078908682666950843","text":"@mindragon I had been just using 1/mo at a time since things are advancing so rapidly, almost let it expire but then saw K3 on the horizon and gave it another month 😍","created_at":"2026-07-19T18:23:48.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":177},"quoted_tweet_id":null},{"id":"2078909638855086109","text":"@leatherytendons They're all gone 😅\nBumped mine up to the $40 so I could get K3\nSeems like 1mo. is about the shelf life of these frontier models 🤷♂️","created_at":"2026-07-19T18:27:36.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":79},"quoted_tweet_id":null},{"id":"2078913720692994532","text":"@Chris_Wozniczek I was literally one click away from cancelling after GLM5.2 dropped too 🙃","created_at":"2026-07-19T18:43:50.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":123},"quoted_tweet_id":null},{"id":"2078913939182698541","text":"@u1tra_instinct same prices, but now they're all sold out. They're indicating these plans will be replaced with new ones, hopefully they grandfather these older ones <3","created_at":"2026-07-19T18:44:42.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":91},"quoted_tweet_id":null},{"id":"2078923285048197406","text":"@DODOREACH definitely! hoping the old ones are better and we get lucky here 😆","created_at":"2026-07-19T19:21:50.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":0,"views":139},"quoted_tweet_id":null},{"id":"2078923670676820384","text":"@Italianclownz The $40 one, bumped it from the $20 the day before K3 so I'd get access 😀\nThe longer ones with discounts seem appealing but given how fast things are progressing it seemed prudent to just use the monthlies!?","created_at":"2026-07-19T19:23:22.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":87},"quoted_tweet_id":null},{"id":"2078938505078509966","text":"@ivanfioravanti I guess this is better than letting too many people sign up and then throttling (looks at GLM-5.2) 🤷♂️","created_at":"2026-07-19T20:22:19.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":145},"quoted_tweet_id":null},{"id":"2078968187815624961","text":"@ToNYD2WiLD One of the few! If you want one now is the time 😁","created_at":"2026-07-19T22:20:15.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":61},"quoted_tweet_id":null},{"id":"2078976587513385437","text":"@cyburke Sooo lucky ngl\nThis was one day from expiring on the $20 plan, caught wind of K3 that same day 🤓","created_at":"2026-07-19T22:53:38.000Z","media":[],"metrics":{"likes":1,"retweets":0,"replies":1,"views":52},"quoted_tweet_id":null},{"id":"2079020026535772190","text":"@sethsaler they must be rolling out the new ones now!","created_at":"2026-07-20T01:46:15.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":18},"quoted_tweet_id":null},{"id":"2079027980089675813","text":"@AnalyticsGenius I don't mind paying $40 for a month to test it!\nDefinitely won't replace my GPT sub <3","created_at":"2026-07-20T02:17:51.000Z","media":[],"metrics":{"likes":0,"retweets":0,"replies":0,"views":19},"quoted_tweet_id":null}]},{"root_id":"2078908327384457584","url":"https://x.com/TheZachMueller/status/2078908327384457584","author":{"username":"TheZachMueller","display_name":"Zach Mueller"},"created_at":"2026-07-19T18:22:24.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2078908327384457584","text":"I’ll just say this on the discourse. If there are no consequences/pushback/something, I’ll probably just become fully OSS/“local” pilled and deal with the dumb-ness of it (I say lightly). \n\nOpenAI has done a great deal (to me) about making me like their brand more, product, etc. \n\nBut if this is actually the way forward, I don’t want the machine god to rule me. I want my helpful buddy to go on adventures with me as I grow and learn throughout life.","created_at":"2026-07-19T18:22:24.000Z","media":[],"metrics":{"likes":21,"retweets":1,"replies":0,"views":988},"quoted_tweet_id":null}]},{"root_id":"2078879331250540881","url":"https://x.com/TheAhmadOsman/status/2078879331250540881","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-19T16:27:10.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2078879331250540881","text":"You wanna learn about Local AI and see it in action?\n\nDO NOT MISS THIS.","created_at":"2026-07-19T16:27:10.000Z","media":[],"metrics":{"likes":291,"retweets":28,"replies":13,"views":30277},"quoted_tweet_id":"2078878406335201738"}]},{"root_id":"2078878406335201738","url":"https://x.com/huggingface/status/2078878406335201738","author":{"username":"huggingface","display_name":"Hugging Face"},"created_at":"2026-07-19T16:23:30.000Z","lang":"en","categories":["local"],"models":[],"posts":[{"id":"2078878406335201738","text":"Join us this Tuesday to learn more on Local AI, from software to hardware 🤗\n\nWe'll be joined by @TheAhmadOsman & @MikeBradleyAI covering hardware setups & local inference with live demo, and @alexocheema & @0xSero on picking model for your hardware, model compression and REAPs 🔥\n\nSet your reminders to not miss out! 🔔","created_at":"2026-07-19T16:23:30.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNmnX00WgAABQXX.jpg:orig"}],"metrics":{"likes":307,"retweets":50,"replies":22,"views":65410},"quoted_tweet_id":null}]},{"root_id":"2078852886151250302","url":"https://x.com/TheAhmadOsman/status/2078852886151250302","author":{"username":"TheAhmadOsman","display_name":"Ahmad"},"created_at":"2026-07-19T14:42:05.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2078852886151250302","text":"An idle GPU is the saddest of stories","created_at":"2026-07-19T14:42:05.000Z","media":[],"metrics":{"likes":198,"retweets":8,"replies":25,"views":8797},"quoted_tweet_id":null}]},{"root_id":"2078830544679243903","url":"https://x.com/aijoey/status/2078830544679243903","author":{"username":"aijoey","display_name":"Joey"},"created_at":"2026-07-19T13:13:19.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2078830544679243903","text":"2.4T parameters is wild.\n\n@Alibaba_Qwen just announced Qwen3.8. It is going open weight soon, and the team says it is one of the strongest models available right now, second only to Fable 5. We will see...\n\nThe Qwen3.8 Max Preview is already live on Alibaba's Token Plan. They also put it in Qoder and QoderWork. \n\nWill this live up to the hype?","created_at":"2026-07-19T13:13:19.000Z","media":[{"type":"photo","url":"https://pbs.twimg.com/media/HNl9XQGWYAAaJ_W.jpg:orig"}],"metrics":{"likes":1,"retweets":0,"replies":0,"views":86},"quoted_tweet_id":null}]},{"root_id":"2078809853993378174","url":"https://x.com/Blackwellboy/status/2078809853993378174","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-19T11:51:06.000Z","lang":"en","categories":["hardware"],"models":[],"posts":[{"id":"2078809853993378174","text":"i will say this again. the marketplace seller who \"doesnt know much about computers\" knows exactly what that 3090 is worth. the one selling a full gaming pc does not. act accordingly","created_at":"2026-07-19T11:51:06.000Z","media":[],"metrics":{"likes":13,"retweets":0,"replies":4,"views":735},"quoted_tweet_id":null}]},{"root_id":"2078799920312787139","url":"https://x.com/LottoLabs/status/2078799920312787139","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-19T11:11:37.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2078799920312787139","text":"If qwen skipped qwen 27b 3.7 release to drop qwen 2.4T 3.8\n\nIt is all forgiven","created_at":"2026-07-19T11:11:37.000Z","media":[],"metrics":{"likes":49,"retweets":0,"replies":8,"views":2556},"quoted_tweet_id":null}]},{"root_id":"2078799559824982140","url":"https://x.com/LottoLabs/status/2078799559824982140","author":{"username":"LottoLabs","display_name":"Lotto"},"created_at":"2026-07-19T11:10:11.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2078799559824982140","text":"Bros maybe just maybe we get some smaller models\n\nQwen 3.8 27b might be real after all\n\nBig release 2.4T parameters\n\nJust skipped 3.7","created_at":"2026-07-19T11:10:11.000Z","media":[],"metrics":{"likes":67,"retweets":0,"replies":6,"views":3506},"quoted_tweet_id":"2078759124914098291"}]},{"root_id":"2078796607798276548","url":"https://x.com/hxiao/status/2078796607798276548","author":{"username":"hxiao","display_name":"Han Xiao ✈️ ICML 2026"},"created_at":"2026-07-19T10:58:28.000Z","lang":"en","categories":["models","performance"],"models":[],"posts":[{"id":"2078796607798276548","text":"super short msg compared to qwen's previous PRs. so they've benchmarked it with confidence and it's at least on par with k3 and upcoming DeepSeekV4 (or not, hence release 3.8 now not later). it shows alibaba's urgency to retake their seat in open-weight models - they've been getting their thunder stolen by GLM and Kimi in last few months. i just had a coffee chat with some insider from ali and he told me there's no trade secret on model training. one team cracks it first, and soon enough everyone else gets there too. pressure on DeepSeekv4 now.","created_at":"2026-07-19T10:58:28.000Z","media":[],"metrics":{"likes":14,"retweets":2,"replies":1,"views":1967},"quoted_tweet_id":"2078759124914098291"}]},{"root_id":"2078788663241298206","url":"https://x.com/sudoingX/status/2078788663241298206","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-19T10:26:53.000Z","lang":"en","categories":["models"],"models":[],"posts":[{"id":"2078788663241298206","text":"@Alibaba_Qwen open weights just claimed number two on the planet. if that holds, the frontier is a download link now, not a subscription, and what exactly is openai and anthropic charging you monthly for.","created_at":"2026-07-19T10:26:53.000Z","media":[],"metrics":{"likes":26,"retweets":0,"replies":2,"views":1664},"quoted_tweet_id":null}]},{"root_id":"2078787862368321753","url":"https://x.com/Blackwellboy/status/2078787862368321753","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-19T10:23:43.000Z","lang":"en","categories":["hardware","models","performance"],"models":["glm-5","glm-5-2"],"posts":[{"id":"2078787862368321753","text":"don’t bother with glm 5.2 on 2 x DGX sparks, it’s a hot waste. Stick with a 4 stack for glm , that runs pretty solid at good tok/s , DSV4 Flash is king of the 2 stack sparks. Step 3.7 pretty solid too, but I find DSV4 much better in actual real world use, gets shit done.","created_at":"2026-07-19T10:23:43.000Z","media":[],"metrics":{"likes":8,"retweets":0,"replies":4,"views":585},"quoted_tweet_id":"2078392353346535670"}]},{"root_id":"2078784689419292848","url":"https://x.com/Blackwellboy/status/2078784689419292848","author":{"username":"Blackwellboy","display_name":"BlackwellBoy"},"created_at":"2026-07-19T10:11:06.000Z","lang":"en","categories":["hardware","local"],"models":[],"posts":[{"id":"2078784689419292848","text":"open source models are about to jump a tier. give it a month, maybe two. local inference boys about to eat good. and the dgx sparks are finally about to show what they were built for","created_at":"2026-07-19T10:11:06.000Z","media":[],"metrics":{"likes":11,"retweets":2,"replies":0,"views":447},"quoted_tweet_id":null}]},{"root_id":"2078777127609208917","url":"https://x.com/sudoingX/status/2078777127609208917","author":{"username":"sudoingX","display_name":"Sudo su"},"created_at":"2026-07-19T09:41:03.000Z","lang":"en","categories":["agents","hardware","local","models","quantization"],"models":[],"posts":[{"id":"2078777127609208917","text":"watch bonsai 3.9gb model crush tool calls. this is running in hermes agent locally, and it doesn't just fire blind, it reasons first before every single action, thinks through what it actually needs, then makes the call. \n\nzero failures the whole run, not one dropped, not one malformed, across a full multi step agent loop. for anyone new here, bonsai is a 1 bit quant of qwen 3.6 27b dense, the exact model i spent all month calling the king of the 3090. \n\n@PrismML took the king, crushed it down to a single sign bit per weight, 3.9gb, small enough to sit next to a gemma 12b, and it still holds an agent loop and reasons before it acts. \n\nsub 4 bit is where models are supposed to turn to mush. someone even clocked it at tool-calling 0 out of 100. not in my runs. this is the warm-up. the real test is the build gauntlet my q4 27b nailed, elements, particle swarm, mandelbrot, and openclaw runs the same one right after. \n\nwe find out if 1bit actually keeps up.","created_at":"2026-07-19T09:41:03.000Z","media":[{"type":"video","url":"https://video.twimg.com/amplify_video/2078774107974610944/vid/avc1/1282x1568/yCbrx0kAaFn3sRCz.mp4?tag=29"}],"metrics":{"likes":202,"retweets":9,"replies":17,"views":16747},"quoted_tweet_id":"2077769809836793893"}]}]} | |
Xet Storage Details
- Size:
- 217 kB
- Xet hash:
- 659d6b815cb4a84a2c55f0af04c5474292c8b1b7615b581fac5023da256a3acd
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.