Introducing @huggingface Sheets: Excel meets AI and unstructured data 📁 Run prompts and models over your data 🌐 Web search for accuracy and real-time information 🎯 Manual edits are used to improve generation 💯 Hundreds of open models and leading inference providers 1/2

Let me know if you try this out, looking for early feedback

Love this 🔥🙌🏽

🔥Direct dataset editing on the Hugging Face Hub🔥 No more downloading & reuploading a 500MB CSV just to fix 3 mislabeled rows. Edit cells, commit changes. Collaborative, fully versioned. Dataset curation finally works like code. https://t.co/HXfES1OCyy

We're launching the world's biggest AI virtual hackathon, with more than 6,000 registrations (and >$1 million in prizes and credits) Join us at the kickoff event in 30 minutes 🔥

Update: Dojo now supports the OpenEnv spec.

We’re launching SWE-fficiency to eval whether LMs can speed up real GitHub repos on real workloads! ⏱️ 498 optimization tasks across 9 data-science, ML, and HPC repos — each with a real workload to speed up. Existing agents struggle to match expert level optimizations!

🔥 IT'S OUT 🔥 Struggling to find benchmarks? Explore our repository for THOUSANDS of easy-to-run, well-documented tasks 🤩 📏 Creators, add yours for more visibility 🔥 Users, find and run models effortlessly with lighteval

Open models are reaching the frontier at an impressive pace. More choices mean more decisions: which model? Local or managed? Which provider? Cheaper or faster? As the offer grows, so does the need for better testing. That's why I built a new integration with Inspect AI for

5/You can also compare :fastest vs :cheapest providers, write custom scorers, and test models in agentic scenarios. All without setting up infrastructure. Drop a comment if you'd like me to run an eval you have in mind on the latest open models.

4/ Evaluate vision-language models with custom datasets I even built a "chihuahua or muffin" style dataset to test VLMs. Results: Qwen3-VL-8B: 0.7 accuracy Qwen3-VL-30B-Thinking: 0.9 accuracy Still unsolved 😂 ~30 lines of Python for custom evals.

3/ Compare the same model across providers I ran gpt-oss-120b across 10 providers on the same task. Scores ranged from 0.80 to 0.84. Hardware and inference implementations vary. Testing helps you find what actually works best for your use case.

2/ Benchmark multiple models on your task Pick models from the Hub, run them in parallel. No GPUs, no downloads—just: inspect eval-set https://t.co/e7703eAxCz --model \ "hf-inference-providers/MiniMaxAI/MiniMax-M2,\ hf-inference-providers/openai/gpt-oss-120b"

Open models are reaching the frontier at an impressive pace. More choices mean more decisions: which model? Local or managed? Which provider? Cheaper or faster? As the offer grows, so does the need for better testing. That's why I built a new integration with Inspect AI for

Unsurprisingly, Kimi K2 Thinking is already number one trending on HF. The AI frontier is open-source!

Fastest and easiest way to download a repo from the HF Hub: $ uvx hf download mlx-community/granite-4.0-h-1b-8bit That's it. Same for uploads, no excuse not to upload your datasets and models. Under the hood: Xet, parallel downloads, huggingface_hub, and of course uv.

The AI frontier is open-source!

you can now push, run, and pull agentic environments from the hugging face hub as spaces. the workflow looks like this: - do `openenv init` to start a new environment from a template - build/port your rl environment - use `openenv push` to push the environment to hub. - from

Has anyone played with Petri (by @AnthropicAI) to test open model behaviours? Would love to chat

Or ... you could just host them on https://t.co/aeVYPxcibJ

If you're interested in agentic RL training, don't miss this awesome guide 👇