moonshotai/Kimi-K3
Image-Text-to-Text ā¢ 2.8T ā¢ Updated ā¢ 2.85k ā¢ ā¢ 6.09k
s
Tom-Neverwinter
AI & ML interests
Making improvements to help the world.
Recent Activity
liked a model 1 day ago
moonshotai/Kimi-K3Organizations
None yet
liked a
model 1 day ago
reacted to badaoui's post with šš 3 days ago
Post
2211
432 GB of ultra-fast HBM4 and up to 23.3 TB/s of memory bandwidth on a single GPU š¤Æ.
Two weeks ago, we got early access to AMD's new Instinct MI455X, and our first goal was simple: make sure š¤ Transformers works on day one.
Over the past few weeks, we worked closely with the AMD team to validate the platform, enable Flash Attention, add torchcodec support for multimodal models, and resolve issues uncovered during testing.
The result:
ā 99.5% success rate across our 24 core Transformers model architectures - already on par with our daily CI on previous AMD and NVIDIA platforms.
The hardware is just as exciting. With 432 GB of HBM per GPU, our early capacity experiments showed more than 3Ć the concurrent long-context requests compared to MI300, thanks to the much larger KV cache capacity.
A huge thanks to the AMD team for the early access and the great collaboration!
Read the full blog š
https://huggingface.co/blog/badaoui/transformers-on-amd-mi455
Two weeks ago, we got early access to AMD's new Instinct MI455X, and our first goal was simple: make sure š¤ Transformers works on day one.
Over the past few weeks, we worked closely with the AMD team to validate the platform, enable Flash Attention, add torchcodec support for multimodal models, and resolve issues uncovered during testing.
The result:
ā 99.5% success rate across our 24 core Transformers model architectures - already on par with our daily CI on previous AMD and NVIDIA platforms.
The hardware is just as exciting. With 432 GB of HBM per GPU, our early capacity experiments showed more than 3Ć the concurrent long-context requests compared to MI300, thanks to the much larger KV cache capacity.
A huge thanks to the AMD team for the early access and the great collaboration!
Read the full blog š
https://huggingface.co/blog/badaoui/transformers-on-amd-mi455
- 1 reply
MOE or changing to A3B
2
#43 opened 6 days ago
by Tom-Neverwinter
reacted to sequelbox's post with šš„ā¤ļø 11 days ago
Post
3512
Multiple NEW RELEASES for Gemma 4 12B!
- Esper 4, our flagship agentic coder: specialist in coding, architecture, DevOps, and MLOps!
- Tachibana-Agent, trained only on code for dedicated, predictable deployment!
- Guardpoint, our structured medical reasoning model: medical diagnosis, management, knowledge, and understanding in structured, concise form!
GET OUR NEW MODELS:
ValiantLabs/gemma-4-12B-it-Esper4
sequelbox/gemma-4-12B-it-Tachibana-Agent
ValiantLabs/gemma-4-12B-it-Guardpoint
Get the datasets for your own training:
sequelbox/Titanium4-DeepSeek-V4-Pro
sequelbox/Mitakihara2-DeepSeek-V4-Pro
sequelbox/Tachibana4-DeepSeek-V4-Pro
sequelbox/Superpotion-DeepSeek-V3.2-Speciale
Esper 4 is also available for Qwen 3.6 27B: ValiantLabs/Qwen3.6-27B-Esper4
We'll be expanding Esper 4 to more models and releasing new models as funding allows - donate for more, faster, better models and datasets: sequelbox/SupportOpenSource
More to come soon!
go build stuff :)
allegra
- Esper 4, our flagship agentic coder: specialist in coding, architecture, DevOps, and MLOps!
- Tachibana-Agent, trained only on code for dedicated, predictable deployment!
- Guardpoint, our structured medical reasoning model: medical diagnosis, management, knowledge, and understanding in structured, concise form!
GET OUR NEW MODELS:
ValiantLabs/gemma-4-12B-it-Esper4
sequelbox/gemma-4-12B-it-Tachibana-Agent
ValiantLabs/gemma-4-12B-it-Guardpoint
Get the datasets for your own training:
sequelbox/Titanium4-DeepSeek-V4-Pro
sequelbox/Mitakihara2-DeepSeek-V4-Pro
sequelbox/Tachibana4-DeepSeek-V4-Pro
sequelbox/Superpotion-DeepSeek-V3.2-Speciale
Esper 4 is also available for Qwen 3.6 27B: ValiantLabs/Qwen3.6-27B-Esper4
We'll be expanding Esper 4 to more models and releasing new models as funding allows - donate for more, faster, better models and datasets: sequelbox/SupportOpenSource
More to come soon!
go build stuff :)
allegra
- 1 reply
any smaller models coming any time soon?
3
#1 opened 13 days ago
by Tom-Neverwinter
benchmaxxed
6
#11 opened 24 days ago
by Tom-Neverwinter
reacted to ginigen-ai's post with šš„ 28 days ago
Post
5206
š³ The RoboCasa Kitchen Leaderboard
What does it take for a robot to handle kitchen chores the way a person does? It has to see (Vision), understand instructions (Language), and actually act (Action) ā and VLA (Vision-Language-Action) models are emerging as the answer. They're the bridge between large multimodal models and real-world embodied control.
RoboCasa Kitchen is a leading robot-learning benchmark in which a single-arm robot (Franka Panda) performs 24 atomic manipulation tasks ā picking up cups and bowls, opening drawers and doors, turning faucets, pressing buttons, and more ā inside a photorealistic simulated kitchen. Because the layout and object placement are randomized every episode, it tests genuine generalization rather than memorized motions. The score (success rate, SR) is the average fraction of the 24 tasks completed as instructed, measured over multiple seeds so results aren't down to luck.
The catch: this benchmark has no official leaderboard, and protocols (number of demonstrations, evaluation setup) differ from paper to paper, leaving scores scattered. Lining the numbers up naively quickly turns into an apples-to-oranges comparison.
This leaderboard fixes that by collecting published scores with their sources and comparing only what is genuinely comparable. It's split into three tables:
š Kitchen 24-task (matched) ā head-to-head under identical conditions (per the RLDX-1 Technical Report). This is the core ranking you can actually trust.
ā Other protocols ā self-reported under different setups (e.g. fewer demos). Not directly comparable, so kept separate.
š¤ GR1-Tabletop ā a different, humanoid-based variant suite, separated to avoid confusion.
Any researcher can submit their own model's score directly, and submissions are reviewed before they appear on the board. Every number links to its source paper, so you can verify it yourself.
š ginigen-ai/robocasa-kitchen-leaderboard
What does it take for a robot to handle kitchen chores the way a person does? It has to see (Vision), understand instructions (Language), and actually act (Action) ā and VLA (Vision-Language-Action) models are emerging as the answer. They're the bridge between large multimodal models and real-world embodied control.
RoboCasa Kitchen is a leading robot-learning benchmark in which a single-arm robot (Franka Panda) performs 24 atomic manipulation tasks ā picking up cups and bowls, opening drawers and doors, turning faucets, pressing buttons, and more ā inside a photorealistic simulated kitchen. Because the layout and object placement are randomized every episode, it tests genuine generalization rather than memorized motions. The score (success rate, SR) is the average fraction of the 24 tasks completed as instructed, measured over multiple seeds so results aren't down to luck.
The catch: this benchmark has no official leaderboard, and protocols (number of demonstrations, evaluation setup) differ from paper to paper, leaving scores scattered. Lining the numbers up naively quickly turns into an apples-to-oranges comparison.
This leaderboard fixes that by collecting published scores with their sources and comparing only what is genuinely comparable. It's split into three tables:
š Kitchen 24-task (matched) ā head-to-head under identical conditions (per the RLDX-1 Technical Report). This is the core ranking you can actually trust.
ā Other protocols ā self-reported under different setups (e.g. fewer demos). Not directly comparable, so kept separate.
š¤ GR1-Tabletop ā a different, humanoid-based variant suite, separated to avoid confusion.
Any researcher can submit their own model's score directly, and submissions are reviewed before they appear on the board. Every number links to its source paper, so you can verify it yourself.
š ginigen-ai/robocasa-kitchen-leaderboard
ungated?
3
#1 opened 2 months ago
by Tom-Neverwinter
reacted to danielhanchen's post with š„ā¤ļø 3 months ago
Post
8942
We made a guide on how to run open LLMs in Claude Code, Codex and OpenClaw.
Use Gemma 4 and Qwen3.6 GGUFs for local agentic coding on 24GB RAM
Run with self-healing tool calls, code execution, web search via the Unsloth API endpoint and llama.cpp
Guide: https://unsloth.ai/docs/basics/api
Use Gemma 4 and Qwen3.6 GGUFs for local agentic coding on 24GB RAM
Run with self-healing tool calls, code execution, web search via the Unsloth API endpoint and llama.cpp
Guide: https://unsloth.ai/docs/basics/api
reacted to SeaWolf-AI's post with š„ 3 months ago
Post
5977
š§¬ Darwin-27B-Opus: 86.9% on GPQA Diamond ā World #5, Zero Training
We are excited to share Darwin-27B-Opus, a 27B model that achieved 86.9% on GPQA Diamond ā ranking #5 globally on the HuggingFace leaderboard ā without a single gradient update.
How? Darwin breeds pretrained models through evolutionary FFN crossbreeding. The father (Qwen3.5-27B) provides the reasoning architecture; the mother (Claude 4.6 Opus Reasoning Distilled) contributes structured chain-of-thought knowledge. CMA-ES automatically discovers optimal per-layer blending ratios ā no human tuning required.
The result surpasses the original Qwen3.5-27B (85.5%), GLM-5.1 (744B, 86.2%), and Qwen3.5-122B (86.6%). A 27B model outperforming 744B ā with zero training, zero data, one GPU, ~2 hours.
We also confirmed hybrid vigor on Korean benchmarks: Darwin-27B-KR (2nd generation offspring) surpassed both parents on CLIcK, winning 7 out of 11 categories. The evolutionary optimizer independently assigned 93% of FFN from the Korean-specialized mother while preserving 93% of attention from the reasoning-specialized father ā autonomously validating our core principle: FFN carries knowledge, Attention carries reasoning.
š Public release: 10 days ā 300+ community derivatives, 120K+ downloads.
š Links:
Darwin-27B-Opus: FINAL-Bench/Darwin-27B-Opus
article: https://huggingface.co/blog/FINAL-Bench/darwin-gpqa
Darwin Family Collection: https://huggingface.co/collections/FINAL-Bench/darwin-family
If foundation models are raw ore, Darwin is the forge. We are just getting started. š„
We are excited to share Darwin-27B-Opus, a 27B model that achieved 86.9% on GPQA Diamond ā ranking #5 globally on the HuggingFace leaderboard ā without a single gradient update.
How? Darwin breeds pretrained models through evolutionary FFN crossbreeding. The father (Qwen3.5-27B) provides the reasoning architecture; the mother (Claude 4.6 Opus Reasoning Distilled) contributes structured chain-of-thought knowledge. CMA-ES automatically discovers optimal per-layer blending ratios ā no human tuning required.
The result surpasses the original Qwen3.5-27B (85.5%), GLM-5.1 (744B, 86.2%), and Qwen3.5-122B (86.6%). A 27B model outperforming 744B ā with zero training, zero data, one GPU, ~2 hours.
We also confirmed hybrid vigor on Korean benchmarks: Darwin-27B-KR (2nd generation offspring) surpassed both parents on CLIcK, winning 7 out of 11 categories. The evolutionary optimizer independently assigned 93% of FFN from the Korean-specialized mother while preserving 93% of attention from the reasoning-specialized father ā autonomously validating our core principle: FFN carries knowledge, Attention carries reasoning.
š Public release: 10 days ā 300+ community derivatives, 120K+ downloads.
š Links:
Darwin-27B-Opus: FINAL-Bench/Darwin-27B-Opus
article: https://huggingface.co/blog/FINAL-Bench/darwin-gpqa
Darwin Family Collection: https://huggingface.co/collections/FINAL-Bench/darwin-family
If foundation models are raw ore, Darwin is the forge. We are just getting started. š„
reacted to DavidAU's post with š 5 months ago
Post
6838
Gemma 3 (1b, 4b, 12b and 27b) - Uncensored full Reasoning/Thinking models fine tuned using top distill datasets.
20 Gemma 3 models 1B, 4B, 12B and 27B with full reasoning using GLM 4.7 Flash, GPT, Claude and Gemini datasets and more fully fine tuned using Unsloth.
Most models are Heretic'ed (uncensored) first, and tuned second.
This vastly improves the model.
Models are also bench marked and in almost all cases exceed org model metrics - and in some cases by a lot.
Enjoy the freedom and more powerful THINKING/REASONING and UNCENSORED Gemma 3s !
https://huggingface.co/collections/DavidAU/gemma-3-reasoning-thinking-models-incl-uncensored
UPDATE: Benchmarks added for almost all models, including "VS" with Heretic (untuned) source models too.
20 Gemma 3 models 1B, 4B, 12B and 27B with full reasoning using GLM 4.7 Flash, GPT, Claude and Gemini datasets and more fully fine tuned using Unsloth.
Most models are Heretic'ed (uncensored) first, and tuned second.
This vastly improves the model.
Models are also bench marked and in almost all cases exceed org model metrics - and in some cases by a lot.
Enjoy the freedom and more powerful THINKING/REASONING and UNCENSORED Gemma 3s !
https://huggingface.co/collections/DavidAU/gemma-3-reasoning-thinking-models-incl-uncensored
UPDATE: Benchmarks added for almost all models, including "VS" with Heretic (untuned) source models too.
Benchmaxxed
š§ š 10
2
#12 opened 7 months ago
by Tom-Neverwinter
Safety or the joke there in
š„ 1
4
#7 opened 8 months ago
by Tom-Neverwinter
Benchmaxxed to the max
1
#1 opened 7 months ago
by Tom-Neverwinter
Benchmaxxed
#2 opened 7 months ago
by Tom-Neverwinter