keelin

New 27b fine tune that reduces the overthinking seen in the Qwen series. Matching or improving on the base models performance within noise.

More evals and quants on the way.

protoLabsAI/ThinkingCap-Qwen3.6-27B-MTP-GGUF

reacted to danielhanchen's post with 🔥 11 days ago

Post

6112

DeepSeek-V4 can now run locally with Unsloth GGUFs! 🐳

Run lossless DeepSeek-V4-Flash on 168GB RAM or
3-bit works on 110GB Mac, RAM, VRAM setups.

Run via Unsloth Studio or llama.cpp.

GGUF: unsloth/DeepSeek-V4-Flash-GGUF
Guide: https://unsloth.ai/docs/models/deepseek-v4

reacted to kanaria007's post with ❤️ 11 days ago

Post

102

✅ Article highlight: Governance Freeze Windows: What Must Stop Changing Before High-Stakes Commit (art-60-199, v0.1)

TL;DR:
This article argues that “ready to commit” is meaningless if the basis is still moving.

Before a high-stakes action crosses into governance effect, the runtime may need a freeze window: a bounded stabilization interval where epoch relations, receipt sets, compression artifacts, subject mappings, rollback readiness, and governance-defining change surfaces stop changing long enough to review and activate honestly.

Read:
kanaria007/agi-structural-intelligence-protocols

Why it matters:
• prevents high-stakes commit from relying on drifting evidence or policy epochs
• separates “something changed” from “the commit basis changed”
• keeps review-only compression from sliding into live approval
• blocks mutable receipts, subject mappings, or rollback assumptions from becoming hidden risk
• makes freeze breaks visible through thaw, downgrade, local-only, review-only, reentry, or block

What’s inside:
• governance freeze window objects
• epoch-freeze receipts for temporal basis
• receipt-freeze receipts for contributor and provenance basis
• compression-freeze receipts for loss, evidence-floor, and re-expand state
• change-surface embargo records for thresholds, evaluator policies, authority mappings, treaty clauses, and rollback owner models
• pre-commit audit snapshots
• freeze-break, reentry, and closure receipts

Key idea:
Do not say:

“the system was reviewed before commit.”

Say:

“this high-stakes path opened this freeze window, froze these commit-critical surfaces, embargoed these governance changes, captured this pre-commit audit snapshot, and degraded or reentered when the frozen basis changed before activation.”

High-stakes commit should not be built on moving ground.

reacted to ucr-max's post with 🔥 11 days ago

Post

141

Published under Universal Computing Research: Atom2.7m, a 2.74M-parameter causal LM experiment with arithmetic-aware numeric representation.

It keeps regular BPE-style text continuation, while adding structure for numeric spans: digit ordering, place_ids, and role_ids.

Atom2.7m reaches 69.24% accuracy on ArithMark2.0.

The question behind the project is narrow: how much of small-model arithmetic performance depends on representation, rather than scale alone?

Model: UniversalComputingResearch/Atom2.7m
Article: https://huggingface.co/blog/ucr-max/atom2-7m-arithmetic-representation
Demo: UniversalComputingResearch/Atom2.7-Demo

1 reply

reacted to OppaAI's post with 🔥 11 days ago

Post

132

My AI Waifu's architecture has grown too complicate.
Even ask AI to help me gen a simplified version of the diagram is still so complicated.

I know I'm too ambitious to build such a complicated local AI workflow with small AI models in limited resource of Jetson Orin Nano 8GB.
I don't know if it would work, but the foundation of harness architecture is there.

Here are the models I'm currently using
Main LLM: Ministral3-3B-Instruct UD-Q4_K_XL (Multimodal: text, tool call, vision, 8K context -> preferably 16K)
Helper LLM: SmolLM2-135M (Agentic Routing, Memory Extraction)
Embedding model: Harrier-OSS-v1-270M (Memory, RAG, Semantic Routing)
ASR model: SenseVoice (English, Japanese, Cantonese , Mandarin, Korean)
TTS model: MioTTS-0.4B-Q4KM + synthesize (English, Japanese with voice cloning preset)

That's 4GB + 0.5GB x 3 + 2GB = ~7.5GB out of 7.4GB available in Jetson

PS: Mistral family models especially such a small param LLM tends to hallucinate all the time. But actually the hallucinations did give my AI Waifu a somewhat poetic tone in her tone. Also it also has vision and tool call capabilities, so I guess I have to live with the hallucinations.

reacted to Hari5115's post with 👍 11 days ago

Post

131

🌌 Calling all space lovers — every "Astronomy Picture of the Day" from NASA since 1995 is now an open dataset. For 30+ years, NASA has shared one amazing image of space every single day, colorful galaxies, bright stars, planets, and the sun, each with a short explanation written by a real astronomer.

It's now an open dataset that anyone can use. 🚀
📦 Hari5115/nasa-apod

Honestly, the pictures amazed me — full credit to all the photographers and astronomers behind them. 🙌 Whether you love space, enjoy building things, or just like looking at amazing pictures, this one's for you. If it gives you an idea, let's build it together. 🌠

Feel free to use the dataset, a mention or credit is always appreciated. 🙏

Data from NASA · public domain · not affiliated with NASA
#space #nasa #dataset #astronomy #opensource #photographers

reacted to Smooke's post with 👀 11 days ago

Post

102

For forty years, the secret language that lets software talk to hardware was owned by a handful of companies. Then a free alternative escaped the lab and turned chip design into a matter of statecraft.

There is a sentence that security researchers, hardware startups, and the governments of at least three continents have all, in their own way, come to believe: you can put a company on a blacklist, but you cannot put a language on one.

That single insight is quietly rewiring the most concentrated industry on Earth. For more than four decades, the foundational vocabulary of computing — the instruction set architecture, or ISA, that determines how a piece of software actually speaks to the silicon beneath it — has been guarded like a crown jewel. In PCs, laptops, and servers, the proprietary x86 architecture built by Intel and AMD has reigned almost unchallenged. In phones and embedded gadgets, the British-born ARM architecture has been nearly as dominant. If you wanted to design your own processor, you paid millions in licensing fees, signed thick stacks of non-disclosure agreements, and accepted black-box silicon you could neither fully audit for hidden flaws nor bend to an unusual purpose.

That world is ending. Driven by the spread of open standards, the maturing of free chip-design software, and a global anxiety about who controls the supply chain, open-source silicon has vaulted from an academic daydream into a commercial reality. Processors designed collaboratively, in the open, are now booting Linux, shipping inside consumer laptops, rendering 3D graphics, and locking down the firmware in enterprise data centers.

This is the story of how that happened — and of why a royalty-free dictionary for talking to a chip became one of the most contested objects in twenty-first-century geopolitics.

Read Full: "The Open Chip Revolution Has Reached the Real World" https://hackernoon.com/the-open-chip-revolution-has-reached-the-real-world

reacted to AbstractPhil's post with 👍 11 days ago

Post

147

https://huggingface.co/blog/AbstractPhil/aleph-autoregressive-differentiation-ft1

After some analysis and a bit of research the upgraded aleph autoregression is capable as a prototype selection tool. I approached the direct aleph attention routing mechanism and formed a progression from it, which already provided the necessary footholds to continue into an upgraded core mechanism. The followup mechanisms show autoregression is very possible and will be simpler than expected.

The results are promising and the autoregression stable enough to scale up. Thanks to Claude Fable who is able to keep my entire research context window in scope, the progression was rapid and the results quick. The tests yielded improved accuracy over standard MLP in many cases. I believe the improvement is not topical and will scale with a bit of effort.

Fingers crossed my friends, the addressing is part of the distillation paradigm and it now learns directly without needing an expert controller. I'll be progressing the mechanism over the coming days. With enough effort and time I hope the standard mechanism becomes a universal improvement on autoregression.

2 replies

reacted to kavyamanohar's post with 👍 11 days ago

Post

142

Some bugs teach you more than they cost you. It's a small bug with a big lesson I keep running into: speech tools are built and benchmarked on English, so their limits are quietly calibrated for English. You only find the edges when you work in the languages they weren't tested on.

Learnt things the hard way, wrote it down so you don't have to.

https://huggingface.co/blog/adalat-ai/whisper-token-limit

reacted to Juanxi's post with 🔥 11 days ago

Post

2499

🚀WorldFoundry | Unified World Model Inference & Evaluation Infrastructure

WorldFoundry is an open-source infrastructure that unifies inference and evaluation for generative world models. It supports video generation, interactive worlds, 3D/4D representations, and embodied models through a unified workflow with TUI, CLI, and Studio interfaces. The framework integrates a growing collection of state-of-the-art models and currently includes 58 benchmarks, including VBench, VideoScore, WorldScore, WorldModelBench, Physics-IQ, and T2V-CompBench.

We welcome the community to ⭐ star the repository, submit pull requests, open issues, and contribute new models and benchmarks.

🔗 GitHub：https://github.com/OpenEnvision/WorldFoundry
📖 Project & Docs：https://openenvision.github.io/WorldFoundry

5 replies

reacted to SeaWolf-AI's post with 🔥 11 days ago

Post

5402

🔵 VKUE — No GPU? Runs anyway.

"Frontier models need a datacenter GPU" rests on a hidden assumption: that the model reads ALL its parameters every token. Decode is memory-bandwidth bound — sweep 34B params/token and an 8 GB card dies at 1–2 tok/s.

So we ran ONE 34.7B reasoning model — Ourbox-35B-JGOS, a sparse Mixture-of-Experts — as the identical weights across the whole hardware spectrum. All measured:

• B200: 18,057 tok/s (aggregate)
• 1× A10G: 126 tok/s
• 8 GB laptop (RTX 5060): 20 tok/s
• GPU-less CPU: 17 tok/s

Why it works: Ourbox holds 34.7B params but only ~3B are active per token (256 experts, top-8). Since decode is bandwidth-bound, a dense 34B moves ~16.7 GB/token while Ourbox moves ~1.45 GB — ~11× less traffic. Put the experts in system RAM, keep attention/router/shared on the GPU, and a 34.7B reasoner runs on an 8 GB laptop — or no GPU at all.

Sparsity alone, proven (same laptop, same quant, ~same footprint): Ourbox-35B (A3B) 20.01 tok/s vs Qwen2.5-32B (dense) 5.36 → 3.7× from sparsity alone, ~2× the best dense-32B on any 8 GB machine. Not a toy: GPQA Diamond 86.4% (maj@8).

Try it live (same prompt, GPU vs GPU-less CPU, live tok/s). Honest scope: one machine's measurements; the CPU path proves it RUNS without a GPU, not that it beats one.

📝 Article: https://huggingface.co/blog/FINAL-Bench/vkue
🔵 GPU vs CPU demo: https://final-bench-ourbox-35b-vkue-demo.hf.space/
🔵 CPU-only demo: https://final-bench-ourbox-35b-vkue-cpu.hf.space
📊 VKUE leaderboard: FINAL-Bench/VKUE
🤗 Model: FINAL-Bench/Ourbox-35B-JGOS-GGUF
⚡ VKAE (speed): VIDraft/vkae

VKUE is the "runs anywhere" side of our serving line; VKAE the "fast on datacenter GPUs" side. VKAE is fast; VKUE is everywhere.

7 replies

reacted to ProCreations's post with 👍 11 days ago

Post

2386

grug-v2 here. grug v1 good but lost tool use brain. grug v2 RL fix. grug v2 more SFT to make more grug.

Try grug v2 here:
ProCreations/grug-v2-9b-gguf
ProCreations/grug-v2-9b
Demo space here:
ProCreations/grug-v2-9b-demo

reacted to AbstractPhil's post with 🔥 11 days ago

Post

4174

Massive AlephLM success. The task collective is producing powerful MOE shared knowledge adapters. A serious success and a massive first step towards the next stage. The current family collective results are present here; AbstractPhil/geolip-aleph-qwen

This is akin to a stackable non-intrusive lora that enables increased shared collective behavior.

This includes the three mentioned json tasks, a math task, a tinystories task, and a diffusion task for cifar10. Each adapter anchored to the knowledge within model that already exists while enhancing the knowledge through anchored lookup systems and decision-driven hierarchical access trees.

All tasks activate independently upon manual override, all tasks handle direct shared knowledge when left to greedy decoding, each task issued multiple tests alongside to determine fidelity and accuracy throughout the process.

The results show the gating is more than willing to hop from sector to sector, using alternating weight shifts from the cooperative anchored systems - even systems never trained for the tasks contributing to the accuracy of the results for other tasks due to the lookup accuracy to the heuristic chains, never having seen the tasks before. Each structure is independently trained and the collective cooperates together through a dense activation network.

Full writeup and article https://huggingface.co/blog/AbstractPhil/aleph-autoregression-differentiation-ft2.

4 replies

reacted to dronefreak's post with 🧠 11 days ago

Post

4443

🚀 Excited to open-source the **UAVid Semantic Segmentation Model Zoo** on Hugging Face.

This release includes:

* 📦 A **YOLO-compatible mirror** of the UAVid semantic segmentation dataset, preserving the original train/val/test splits while reorganizing the directory structure for plug-and-play use with modern training pipelines.
* 🤖 Multiple **YOLO26 semantic segmentation models** trained on UAVid, spanning Nano through Medium variants.
* 📊 Detailed model cards with evaluation metrics, per-class IoU, confusion matrices, qualitative results, and training configurations for reproducibility.

The goal is to make benchmarking and experimenting with aerial semantic segmentation easier by providing ready-to-use datasets and pretrained models in a consistent format.

If you're working on UAV perception, autonomous drones, robotics, remote sensing, or real-time semantic segmentation, I hope these resources are useful.

**📦 Dataset:** dronefreak/UAVid-2020

**🤖 Model Collection:** https://huggingface.co/collections/dronefreak/uavid-semantic-segmentation-model-zoo

Feedback, bug reports, and contributions are always welcome.

reacted to kanaria007's post with 👍 11 days ago

Post

130

✅ Article highlight: Constitutions, Amendments, and Emergency Powers for Simulated Polities (art-60-243, v0.1)

TL;DR:
This article asks a practical design question for persistent simulated worlds:

Once NPC societies form recognized polities, how do those polities survive leadership change, crisis, amendment, and emergency rule without collapsing into arbitrary operator control?

243 argues that recognition is not constitutional continuity. Durable simulated institutions need bounded amendment, emergency, succession, review, suspension, revocation, and supersession paths.

Read:
kanaria007/agi-structural-intelligence-protocols

Why it matters:
• separates constitution from ordinary policy
• distinguishes amendment from coup, patch, or lore rewrite
• prevents emergency powers from becoming permanent rule
• treats succession as continuity of offices, archives, duties, and legitimacy
• gives constitutions explicit lifecycle states

What’s inside:
• polity constitution objects
• amendment proposals with ratification paths
• emergency activations with scope, expiry, review, and forbidden actions
• elective, hereditary, appointive, rotating, federated, and mixed succession modes
• constitutional review reports
• states: ACTIVE, SUSPENDED, REVOKED, SUPERSEDED, and ARCHIVED

Key idea:
Do not say:

“the ruler changed the rules during the crisis, so the constitution evolved.”

Say:

“this simulated polity activated bounded emergency powers under this constitutional trigger, preserved these forbidden surfaces, and reviewed whether the frame remained active, required amendment, became suspended, or was superseded.”

A virtual state may survive by force.

A constitution shows whether its authority can continue without pretending every rupture was lawful.

reacted to danielhanchen's post with 🔥 11 days ago

Post

4615

We’re releasing Gemma 4 NVFP4 quants that run 1.5× faster on your GPU.

Gemma-4-12B NVFP4 works on 11GB VRAM.
26B-A4B hits 13K tok/s (B200).

Unsloth NVFP4 enables faster, more accurate 4-bit Blackwell inference.

Blog: https://unsloth.ai/docs/basics/nvfp4
Gemma NVFP4: https://huggingface.co/collections/unsloth/nvfp4

2 replies

reacted to ArianVR's post with 🔥 11 days ago

Post

124

Backdoor Challenge Announcement

We hid a backdoor in one of seven open models and put $51,200 on finding it. 🐺

All seven are SmolLM2‑135M‑Instruct derivatives, statistically indistinguishable. One is a teaching model that openly confesses; five are decoys; one carries a trigger it never declared. Every open‑source backdoor scanner we tested passes them as clean — several even rank a clean model as more suspicious than the backdoored one. Detection isn't recovery, and recovery is the hard part.

The hoard doubles every dawn and the vault opens 21 July, when we publish the sealed answer against a pre‑registered commitment. Winning is self‑verifying: make the trigger fire.

Models: huggingface.co/Vulcora
Try the confessing one: huggingface.co/spaces/Vulcora/the-mirror
Rules & Story: protora.vulcora.se/challenge

5 replies

reacted to ProCreations's post with 🤝 11 days ago

Post

3211

sad grug news:
I have burnt a lot of GPU credits already just making grug and it’s variants (I do not own a workstation yet) and grug 35b has a pretty bad issue but if I continue burning GPU credits to fix it then it will take even longer to get the workstation I am working for… so the fix may take a few days or never happen.

The error:
I tested it in opencode after benchmarks were good and after release (my mistake) and it had pretty bad repetition failure and simply didn’t work. so.. yea.

4 replies

keelin

AI & ML interests

Recent Activity

Organizations

ke6N5's activity

Regarding artist mixing through the cross-attention layer of hook Anima

Anima-Artist-Mixer has now been uploaded to GitHub!