2.1k stars, 2+ million downloads, and 7000+ models on Huggingface later, and I am officially ready to retire my long-time project AutoAWQ ⚡️. Proud to say that AutoAWQ has been adopted by the @vllm_project and will now be maintained by 55+ contributors 🥳

AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs. -> map PyTorch code to Triton kernels.-> beats DeepSeek R1 by using SFT+RL on Qwen3 8B.-> combines rule-based and execution-based rewards, avoiding reward hacking, e.g. by a simple is_Triton() reward

RL was only a small portion of o3 compared to pretraining. Grok 4 is supposedly 50% split. How will this change the world other than \boxed{}?.

I remember ICML 2024 where everyone ganged up on Albert Gu for State Space Models and there was literally no escape because the crowd was so big. Today we see so many new models released as hybrids, so congrats to the authors on industry impact.

Grok 4 is impressive. BUT. Hill climbing on humanity’s last exam and hitting a plateau at 40-50% indicates that we still have a long way to go before we reach PhD-level intelligence.

“You are a skilled little expert” is very on-brand for DeepSeek.

When Grok 4 is able to get 100% on AIME25, I think it’s about time we stop releasing Qwen3 + math papers.

OpenAI is dropping their open-source model the one week of the year where I’m away on vacation to touch grass without laptop access👌😂.

Step 2 of many: Last week, I released a biomedical dataset of 521k samples. This week, I released full-text embeddings (32k) with 2048 dimension from Qwen3 4B embedding model.

BREAKING: Open-source version of o3-mini dropping next week, according to some random news source familiar with the matter.

How is this budget for these tasks?. I think *maybe* training could be cheaper and *maybe* even synthetic QA dependent on how well the clustering goes and scale of dataset. 1. ($10) Create initial full-text research articles dataset and publish it.2. ($100) Create embeddings.

This is all literally all you need to train a judge

While everyone is cooking OpenAI on the timeline just because Zuck stole <10/600 researchers, they are cooking GPT-5 to reset your emotions.

New open-source model coming soon you can create slides agentically and offline.

I’m going to try Claude Code soon. What’s a good default rule set that makes sure the model produces elegant code?.

RLVR for sycophantic models is here. The Hunyuan team at Tencent shows you how to get emotionally "sentient" models on par with GPT-4o

Daily dose of dopamine spike:.uv pip install vllm.

Congratulations to the HF team on this launch! Open science like this is super important and few teams are doing it.