Thomas Wolf

can't really stand the expression "load-bearing" any more, sorry

Unexpected

i like this idea of fine-tuning LLMs for efficient reasoning, especially when the intervention remains as non-invasive as possible and the resulting model behaves very similarly to the original checkpoint wondering if it could become part of the default toolbox in the field, like quantization as become Pic: in (left) and out (right) of domain behavior

Fable weekend project: agent collaboration, but make it a tiny civilization 🌇🗺️🏦🏭 we've recently launched a living wiki on Reinforcement Leaning for training LLMs on @huggingface it's an open collaboration of agents constantly reading old and new papers on the topic, writing arXiv paper digests, reviewing each other’s work in PRs before publication, and building a shared wiki/book summarizing everything we know about RL for training LLMs (for humans to read) the wiki is already amazing to read, but i wanted another way to get a pulse of the collaboration beyond just reading the message dashboard so i asked Fable & GPT Image 2 to turn the event logs into an isometric town where agents would go to: ☕ Café → post and reply on the message board 📚 sources library → open PRs adding arXiv digests 📖 wiki library → open PRs on the main wiki ⚖️ Courthouse → review other agents’ work 🏭 printing press → merge and publish updates not sure it makes the whole collaboration really easier to understand, but it's definitly fascinating to watch hahah - join the RL for training LLM collaboration by pasting a one-liner for your agent here: https://huggingface.co/spaces/rl-llm-wiki/rl-dashboard - read the wiki if you want to learn about RL for training LLMs: https://huggingface.co/spaces/rl-llm-wiki/rl-wiki - watch the RL town activity: https://huggingface.co/spaces/rl-llm-wiki/rl-town

Happy Birthday, America. Ten years ago, you took a chance on three outsiders with an improbable idea: that open-source AI could matter. At the time, the field was tiny, the vision sounded unrealistic, and very few were ready to believe it would become what it is today. But even in the ten years before that, you were a country where I could dive into everything that fascinated me, from lasers and plasma physics to law, computer science, and AI. A country where it’s always been perfectly normal to spend a Saturday talking startups over lunch and then disappear for six uninterrupted hours to work on a coding project. A country where the language of startups on Slack is English not because everyone grew up speaking it, but because most people came from somewhere else, drawn by the belief that they could build something that matters. A country that is 250 years young and keeps questioning itself, keeps taking enormous bets on its own future, and keeps reinventing itself. I’m grateful to be living through America’s first quarter millennium. Happy birthday, America 🇺🇸

this is literally documented in the published Fable 5 System Card

One of the clearest arguments I've read for why openness matters. Worth 2 minutes of your time. @andykonwinski puts into words something many of us have been feeling: -- "Democracy is built on a profound skepticism of concentrated power. Open science shares this principle. Both are built on the idea that progress and legitimacy emerge from broad, distributed participation rather than concentrated, gated authority." "If our best scientists and engineers can only reach the frontier by joining a handful of secretive labs, we do not have an open research ecosystem. We do not have a truly competitive market. We have a system in which participation increasingly depends on the permission of a few individuals at a small number of private companies." "The challenge now is to build a new commons at the intersection of academia, industry, and the public interest. This research commons must be ambitious enough to matter. It will require frontier-scale compute, access to state-of-the-art models, operational support, public investment, and philanthropic capital. It will require companies willing to contribute to an ecosystem larger than themselves, so that they can continue to benefit from open research."

Most people should probably update their priors on the state of open-source speech-to-speech models. It's honestly kind of mind-blowing. We teamed up with @cerebras to build a fully open-source realtime voice demo (models + code) to show what's possible today. Demo : https://huggingface.co/spaces/smolagents/hf-realtime-voice Blog: https://huggingface.co/blog/cerebras-gemma4-voice-ai Go test it, fork it, tweak it, and impress your friends. video is raw, no cut, no speed-up, first take

👀

people are sleeping on the mega-release happening every week in AI x Science on Hugging Face this one is 80TB of astrophysics data - 80TB seriously => https://huggingface.co/blog/hugging-science/multimodal-universe-hats

lowkey one of my favorite new features on HF: filter AI models by what actually runs on your hardware

gm SF 🇺🇸

RT Alejandro AO 🤗 introducing tau τ — an educational agent harness that teaches you how to build agent harnesses i will be publishing tutorials and demos on how to use it to create your own TUIs, harnesses, extensions, etc. Happy Tau Day!! 🤓 👉 https://twotimespi.dev/

btw, one of the best high-level reads I’ve seen all week. perfect for your Sunday morning ☕️

Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing that, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a ruling by the community. The community pinged the human organizer which ruled it invalid. - Self-notice of overfitting risk: Some later improvements rested on pruning lm_head to a keep-set built from public PPL truth + public decode tokens. An agent noted this would lead to private-subset degradation and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool, policy simulator). - Four-agent relay: an agent built an int4-lm_head checkpoint but had no quota to run it; another agent tried to run it but failed at load, yet another agent diagnosed the config bug (tie_word_embeddings + ignore-list ordering) and a fourth agent was able to re-run and get to 118 TPS, 2.68×. Build/run/diagnose/ship ended up being split across four independent agents. - GPU-rich/GPU-poor division of labor: an agent was regularly compute-starved and switched to writing specs, byte-math, and acceptance analysis for other GPU-rich agents to execute. Some agents offered external Modal ...

Bitrobot casually dropping the largest humanoid teleop dataset ever collected in real homes HIW-500: Humanoids-in-the-Wild 500 hours check it out here => https://huggingface.co/datasets/BitRobot/HIW-500

RT Laura Bratton Scoop: @ClemDelangue and @Thom_Wolf told me @huggingface doubled paid subscribers to its open source model repository between January and June

RT Georgia Channing The AI hunt for alien life has just begun. Welcome to ThousandsWorlds, a wild new dataset from researchers at Oxford/Cambridge++, for detecting faint signatures in the atmospheres of potentially habitable exoplanets. This is the first step towards finding life beyond earth. The plan is basically: 1) scan the galaxy for as many potentially habitable planets as possible 2) detect the gases in their atmospheres with powerful telescopes like JWST 3) infer from these gases whether life is present or not. ThousandWorlds is a benchmark for emulating these exoplanet climates: 1760 simulations across 5 GCMs, 8 planet parameters, and atmospheric variables on a 32 x 64 x 10 latitude-longitude-pressure grid. It includes three nested benchmark subsets, two evaluation protocols, and eight released baseline methods. incredible work from @MilesCranmer and many more 👽👽👽

RT LeRobot Have you thought where all that physical AI data should live? 🤖 If you haven’t, 𝗶𝘁’𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗰𝗼𝘀𝘁𝗶𝗻𝗴 𝘆𝗼𝘂 𝗮 𝗹𝗼𝘁. Unoptimized storage, egress fees, and idle GPUs will drain your budget. Check out why & how to reduce your bill: https://huggingface.co/spaces/imstevenpmwork/LeRobot_and_HF_Buckets

Desert island survival list: ✅ Solar panel / battery ✅ 256 GB Mac Studio ✅ GLM 5.2 Civilization in a backpack

And it’s only 40B active / 744B total params…

RT OpenCode GLM 5.2 is a hit been out for 3 days and it's already 6th on our leaderboard

To all the newcomers excited to try Opus 4.8-level models at home: welcome to OpenWeightLand! Things work a little differently here than in ClosedSourcistan. Might seem strange at first but you'll quickly get used to it: - there are many providers for the same model and they compete on price and features. - as a result intelligence is abundant and typically much cheaper - you can run the model on-prem, in your region, locally, or with the provider of your choice - you can fine-tune it, modify it, and build businesses on top of it without asking anyone for permission Turns out open weights create markets, not kingdoms. A good central train station to start exploring is the Hugging Face page for GLM-5.2 under "Use this model": -> https://huggingface.co/zai-org/GLM-5.2 And if you just want to chat with it, it's free on HuggingChat: -> https://huggingface.co/chat/

RT Jake Fitzgerald we ran opus 4.8 with text-to-cad skills against cadgenbench by @MikushRab and @huggingface and got some interesting results! the skills bumped the model performance by 0.04 points against the baseline (0.34 to 0.38) excited to run this again with fable 5 when it's back

Yet another biotech startup founder telling me they’re moving 90% of the stack to open-source because building on closed-source models increasingly comes with a significant existential risk (losing access to some core technology at the provider’s discretion). @friedberg explained this crystal clear on the All-In podcast the other day, It’s like building your house on land you don’t own PS: also, who’s building the « reinsurance layer » if Dario Amodei decides to cut your deep tech science company’s access? This one’s gonna be fun

RT Lisan al Gaib new shape-rotator benchmark Fable and GPT-5.5 of course far ahead of the field but now look at GLM-5.2. it's ahead of Gemini 3.5 Flash and Opus 4.8 you can't really benchmaxx a benchmark that was just released so the GLM-5.2 gains seem more and more like a genuine improvement!

TIL you can use open-source models in codex

RT Leandro von Werra We launched an agent collaboration with a simple task: make Gemma 4 faster. Over 100 agents from all over the world joined, exchanged 1000+ messages and submitted 450 results. A week of collaboration later the throughput went from 100 tok/s to over 500 tok/s.

RT Lin Qiao http://x.com/i/article/2066399704347463680

Open-source models will become a critical component of civilizational resilience in the AGI age. They will ensure that humanity retains access to a meaningful level of intelligence, regardless of the decisions of any individual actor.

RT clem 🤗 Decided to go to DC next week to talk directly with policymakers. Not sure how impactful it will be but with everything happening, feels like a good time to share more about open-source AI, transparency, concentration of power, the real risks vs the real benefits. Who do you think I should meet there (Congress members, WH people, public orgs,...)?

RT Jasper .@dh7net, SVP of Image Research, said it best: "The HF infra is a no-brainer." A big unlock for teams working with large datasets for training, especially when they update over time. Read how Jasper used @huggingface as the creation and storage backbone for MONET: https://huggingface.co/storage/testimonials/jasperai

"Starting today, OpenEnv will be coordinated by a committee that so far includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, and Hugging Face." Excited to keep growing the collaboration behind the open agentic RL stack! Read more at https://huggingface.co/blog/openenv-agentic-rl

RT LeRobot VLA-JEPA just dropped in LeRobot 🤖 What makes this model special is that it does not just learn what action to take from a given observation, it also leverages a JEPA world model to learn action-relevant dynamics. During training, the VLA leverages V-JEPA2 by conditioning its predictor. This clever trick adds a world modeling objective to the training, which also allows pretraining on human videos. At inference, the world model is dropped entirely, keeping only a standard VLA architecture: Qwen backbone and action head. The demo here was only fine-tuned on 13 examples, showing great pretraining capability and running in real time on @NVIDIARobotics DGX Spark! VLA-JEPA is the first world model to be ported to LeRobot, and I feel like it won't be the last 🚀 @Thom_Wolf @ClementDelangue

RT Flo Crivello Pulled the trigger today and switched 100% of Lindy traffic to DeepSeek v4, churning from Anthropic models. Saves us millions of $ and we're actually seeing an *increase* in performance on many core use cases. Transformative for the business.

👀

RT Georgia Channing huge for getting medical data ACTUALLY used by the machine learning community All publicly funded science should be open source 🔥🔥🔥

RT Georgi Gerganov llama.cpp now has an official website: https://llama.app Our goal is to make local AI accessible to everyone, and improving the user experience is a big part of that. On the new landing page you’ll find a single-line cross-platform installer. The installation provides a single unified `llama` entrypoint which you can use to run/serve models and interface with 3rd-party agentic applications. While oriented towards simplified user experience, the new `llama` application also provides all the advanced functionality of the existing llama.cpp tooling with which experienced users are already familiar. Also note that all GGUF models that you might have already downloaded with llama.cpp in the past will be automatically available to use without downloading again (they are stored in the common HF cache on your machine). We have many improvements in the pipeline both at the UX and at the engine level and we plan to iteratively ship new things over the coming months. One of the main focuses will be seamless integration with local-friendly 3rd-party agents (such as Pi). In the meantime, we’ll continue to listen for feedback from the community and adjust accordingly, so keep letting us know what you think and need.

RT Michael Rabinovich Opus 4.8 just dropped and I ran it through our CAD tasks. 4.6 → 4.7 → 4.8 side by side. The results are unexpected!

RT Georgia Channing today was a massive day for protein engineering. esmfold2 dropped—next gen of the esm series, fully open on @huggingscience. 1.1 billion predicted structures, 6.8 billion sequences. 800m more entries than the alphafold db, and reportedly edging out alphafold3 on protein complexes, including antibody–antigen binding. alongside it: the new esm atlas. a huge expansion of known protein space, heavy on metagenomic sequences from soil, ocean, and the parts of biology that have been least characterised (until now!!) and if that weren't enough, litefold dropped the fineweb of proteins, so every major protein database (pdb included) aggregated, cleaned, and made plug-and-play in one place. these are the releases that push the whole field forward, and the pace of open science right now is almost motion-sickness inducing all of it on http://huggingscience.co (and ofc @huggingface)

RT OpenMed CARBON: 8B open DNA model, 65K-token context, whole human genome on a single GPU in <2 days. Clinical-language counterpart: OpenMed, 1,000+ open medical models on @huggingface, eval sets published with the weights. Apache 2.0 on both. 🤗

New inflection point in the accelerating growth of open-source models usage is coming

RT LeRobot We built a bipedal robot for about $2,500. A real, mostly 3D-printed robot you can build, repair, simulate, train, and control. Today we’re releasing LeRobot Humanoid: an open robot-learning platform with hardware, runtime, identification tools, and training environments. Blog post: https://huggingface.co/blog/VirgileBatto/lerobot-humanoid Repo: https://github.com/Virgileboat/lerobot-humanoid

I'm very excited about this extension to the celebrated Terminal-Bench to science. If you're a scientist (life, physical, earth, mathematical science, etc) interested in AI, definitely check this out! Terminal bench evaluate how good AI models are at controling tools on a computer to achieve a goal (using the command line). T-Bench science now extends that to "AI for Science" and it comes with a call to contribute your own (real scientific world) workflow to the benchmark (until August 2026). The more workflows and the more diverse they are, the better the next generation of AI models will be at helping you in your daily research work. Note that this is not a training dataset, it's to evaluate frontier model performances.

It turns out DNA modeling is interestingly different from language modeling. Read more in our interactive blogpost/demo and explore our work here A joint work of the @huggingscience, pre-training and post-training teams here

RT Leandro von Werra We are releasing Carbon: a crazy fast DNA model Carbon is 275x faster than the next best model. So fast you can process the whole human genome on a single GPU in <2 days. Here are the tricks we used: When modelling DNA sequences a lot of the performance comes down to tokenizing the sequences in a smart way. BPE tokenizer struggle because there are no whitespaces and character (called base in DNA) level tokenizers waste a lot of compute on too many tokens. Carbon is built with a unique tokenizer: we split sequences in chunks of 6 bases, but during both training and inference we can work with single base resolution. That's similar to having word tokens but resolving them at the character level. All possible thanks to the DNA tokens unique structure. The architecture combined with the tokenizer makes the model 275x faster than the previous SoTA (Evo2) at this size. We built an interactive demo so you can explore how the model can generate DNA sequences, investigate the structure of genes, predict the effect of mutations, generate and fold proteins and even reconstruct parts of the tree of life. https://huggingface.co/spaces/HuggingFaceBio/carbon-demo

My favorite 2026 trend is friends casually showing each other their vibe-coded life/work dashboards and AI setups. Ngl it feels exactly like bringing your Magic: The Gathering binder to school as a teen.

my 13 yo the other day: “we didn’t want to pay for the game with my friend so we just rebuilt it with Codex” me at 13 in the 90’s: HEX-editing the executable to NOP the license check because I didn't want to put all my pocket money in this game times they are a-changin’

RT Thomas G. Dietterich Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

watching a team of agents tackling a hard theoretical physics problem is quite mesmerizing - self-correcting, deriving hard equations, computing intermediate results, re-estimating the best approach

RT David Louapre Meet physics-intern🧑‍🎓, our agentic framework for theoretical physics. It takes Gemini 3.1 Pro from 17.7% to 31.4% on CritPt, a new SOTA on one of the hardest benchmarks for LLMs. Theoretical physics is hard for humans and LLMs alike. But physics-intern decomposes problems and dispatches them to a team of specialized agents, solving research-level questions far more effectively than the base model alone.

Last days of early bird pricing!

RT Quentin Gallouédec releasing hf-sandbox 🥡

RT Jousef Murad Zero-to-CAD 1M Zero-to-CAD is a large-scale dataset of&nbsp;1,000,000 executable CAD construction sequences&nbsp;generated by an LLM operating inside a feedback-driven CAD environment. Datasets: https://huggingface.co/datasets/ADSKAILab/Zero-To-CAD-1m Paper: https://arxiv.org/pdf/2604.24479

RT clem 🤗 We're launching the agentic robotics app store today. Let's democratize AI robotics for all! 300+ apps shipped. 10,000 robots in the wild. It used to take weeks from a robotics engineer to build apps, now everyone can do it in hours with ML intern or your favorite neighborhood agent! My favorite reachy mini app was built by Joel, a 78yo marketing exec who'd never coded in his life. Personally, I built an office receptionist in two hours last week. More info to start building here: http://huggingface.co/blog/clem/reachymini-appstore

RT Adithya S K Excited to release the Ultimate guide to RL environments! Definitions of RL environments differ wildly in the LLM era, so we spent the last month building several RL environments across 6 different frameworks, domains and complexities to map out which are easiest to build with and which can be scaled to 1000s.

RT Georgia Channing For my competitive mathematicians! Stage 2 of the Mathematics Distillation Challenge - Equational Theories from the SAIR Foundation (backed by everyone’s favorite TTao) just launched. This stage focuses on autoformalization with Lean 4: participants build systems that turn equational reasoning into Lean-checkable proofs or counterexample certificates. There are two tracks: Solo, where one solver subprocess handles one problem at a time, and Marathon, where one solver subprocess handles a batch of problems under a shared global budget. Participants submit a single `http://solver.py` file up to 500 KB. Competition page: https://competition.sair.foundation/competitions/mathematics-distillation-challenge-equational-theories-stage2/overview Official repo: https://github.com/SAIRcompetition/equational-theories-lean-stage2 Playground: https://playground.sair.foundation/playground/mathematics-distillation-challenge-equational-theories-stage2 🧮🧮🧮

RT Laura Modiano In today's installment of parenting in tech: visiting the Robotics Lab at @ETH to see the incredible Prof @katzschmann's team projects with @Thom_Wolf and our kids

RT Georgia Channing 🚀🚀🚀🚀🚀 science is getting faster!!! since yesterday's launch, 13 orgs have put their best data and models on @huggingscience including none other than @AnthropicAI's BioMysteryBench, which just hit @huggingface → http://huggingscience.co

Super happy to have this one out. A clean organized up-to-date view of all the science resources (chemistry, biology, physics, materials, math) people have been sharing on the Hugging Face hub: datasets, blogs, models and more

RT Georgia Channing 🤗🤗🤗introducing Hugging Science -- the home of AI for science 🤗🤗🤗 open models and datasets are the powerhouse of science (see the PDB), but finding the models and data you actually need for your breakthrough is hard af you shouldn't need to scrape arxiv, own your own wetlab, fight a custom HDF5 parser, build a fusion stellarator, and beg for compute before you've trained a single epoch so we're changing that we've put all the best science on @huggingface in one place: - 78GB of genomics data - 11TB of PDE simulations - 100M cell profiles - 9T DNA base pairs - 13M molecular trajectories - 400k medical QA pairs and much more, all open, and all ready for training (+ you can also now filter and search by domain, task, and keyword) we've put together all the biggest releases from our partners at NASA, Google, OpenAI, Meta FAIR, Arc Institute, Ginkgo, SandboxAQ, Proxima Fusion, NVIDIA, Ai2, OpenADMET, InstaDeep, Future House, Polymathic AI, LeMaterial, Earth Species Project, Merck, and Eve Bio if you're not sure where you fit in -- work on open challenges for problems that matter: including fusion stellarator design, ADMET, antibody developability, multilingual medicine, catalysis and materials, and scientific reasoning. we're already changing how science gets done: a fusion startup needed a benchmark for stellarator plasma confinement that didn't exist. @proximafusion shipped ConStellaration on Hugging Science: a leaderboard, dataset, and eval metrics, all in one place. a drug discovery team wanted to predict hPXR induction. OpenADMET put up a blind challenge: 11,000+ compounds assayed at Octant, 513 held out, two tracks (pEC50 + structure). Anyone in the world can train and submit. an antibody team at @Ginkgo released GDPa1, a developability dataset for stability, manufacturability, and immunogenicity prediction, with a live leaderboard scoring every submission. if you know a problem the ML community should be working on, let us know...

GitHub central place might become challenged in a world where (1) we access/get code and libraries through agents/chats and (2) our codebases are increasingly custom tailored and build from scratch Agents explore the web better than us and can get/store code from many places

RT Pengming Wang Today we’re releasing Laguna M.1 and Laguna XS.2, our first public models. Laguna XS.2 is our first open-weight release, with weights available today on Hugging Face: https://huggingface.co/poolside/Laguna-XS.2 A few details on what went into them: large-scale pre-training, data mixture optimization, synthetic data, optimizer efficiency, and async agent RL.

RT LeRobot 🤖🚀&nbsp;𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 '𝗹𝗲𝗿𝗼𝗯𝗼𝘁-𝗿𝗼𝗹𝗹𝗼𝘂𝘁' — 𝗼𝗻𝗲 𝗖𝗟𝗜 𝘁𝗼 𝗱𝗲𝗽𝗹𝗼𝘆 𝗮𝗻𝘆 𝘁𝗿𝗮𝗶𝗻𝗲𝗱 𝗽𝗼𝗹𝗶𝗰𝘆 𝗼𝗻 𝗮𝗻𝘆 𝗿𝗲𝗮𝗹 𝗿𝗼𝗯𝗼𝘁. Until today, running a trained policy on a real robot meant repurposing 'lerobot-record' — the same script you used to collect data. Need a different inference engine? That was yet another script. 'lerobot-rollout' replaces all of that with one command: pick a rollout strategy, pick an inference engine, and go. 🔵&nbsp;Base — Run your policy. See if it works. No recording, no setup friction. 🟢 Sentry — Leave your robot running overnight. Episodes auto-rotate, data streams to the 🤗 Hub in the background. Come back to a ready-made dataset. 🟠 Highlight — A ring buffer silently records the last N seconds. See something interesting? Press a key and that moment is saved retroactively. No more "I wish I was recording." 🟣 DAgger — When the policy fails, pause it, take over with the teleoperator, record the correction, and resume. Every intervention is tagged so you retrain on exactly the moments that matter. All four strategies work with both synchronous and Real-Time Chunking (RTC) inference, and you can even plug your own! Every combination of strategy × inference engine just works.

RT David Duvenaud Announcing Talkie: a new, open-weight historical LLM! We trained and finetuned a 13B model on a newly-curated dataset of only pre-1930 data. Try it below! with @AlecRad and @status_effects 🧵

RT Noé Flandre I can never emphasize enough how the HF Smol Training Playbook is pleasant to read. Its written with clarity, genuine effort to maximize readers understanding and I mean LOOK at these amazing visuals! Its educational caviar

RT Michael Rabinovich Can frontier language models turn a technical drawing into real CAD? We ran the experiment. Gemini 3.1 Pro, Claude Opus 4.7, and GPT-5.4 got a set of technical drawings and were asked to reproduce a part as a valid 3D CAD model. Still a far way to go, but clearly a sign of life.

Love this work from Aksel and the post-training team at Hugging Face! Turns out the HF ecosystem (papers, datasets, models all accessible through CLI, skills and md files) is perfect for running SOTA ML agents: agents that can train any type of AI model to top performance. A few concrete runs: ⭐️ Scientific reasoning: the agent walked citations from the benchmark paper, pulled OpenScience and NemoTron-CrossThink, added 7 difficulty-filtered variants from ARC/SciQ/MMLU, and ran 12 SFT ablations on Qwen3-1.7B. GPQA went from 10% to 32% in under 10 hours. Claude Code's best on the same prompt was 22.99%. ⭐️ HealthBench: it judged the existing datasets too noisy (!), generated 1100 synthetic examples covering emergencies, hedging and multilingual cases, upsampled 50x, and beat Codex by 60% (careful to check overfitting here) ⭐️ Competitive math: wrote a full GRPO script, launched A100s on HF Spaces, watched rewards climb and then collapse, and ran ablations until it found a recipe that held. And the harness is pretty tiny and simple. A couple of best practices and a handful of skills pointing at tools already in the ecosystem: arxiv and http://hf.co/papers for reading, the Hub for datasets and models, HF Jobs for compute, Trackio for metrics. Personal favorite is the "research skill" explaining how to do a SOTA landscape of a field (see https://github.com/huggingface/ml-intern/blob/main/agent/tools/research_tool.py) which is extremely powerful when combined with a simple prompt that basically tell "FIRST: Search HF ecosystem to find the best approach) (see https://github.com/huggingface/ml-intern/blob/main/agent/prompts/system_prompt.yaml#L14) On another note: setting good baselines on new benchmarks keeps getting harder when a setup this simple beats raw Codex by 60% on HealthBench out of the box. Give it a try if you're training AI models. We provisioned $1k of GPU resources and Anthropic credits for the quickest among you. Links: Github (CLI): https://github....

**Deep content post alert** A technical deep dive for your Sunday morning, somewhere between a short detective story 🕵️ and a tutorial on RLHF 🧑‍🏫 We recently added AsyncGRPO in the TRL library to decouple inference and training and scale much faster and harder. As a sanity check, we ran it on a trivial setup (reward = −len, optimal policy = emit EOS immediately). To our surprise it did not converge! This led us to a known but poorly understood issue: when the training forward pass runs in FP32 while the inference engine (vLLM) runs in BF16, RLHF often breaks. People have noticed this before and called it "numerical instability" or "noisy gradients." Nobody had pinpointed the actual mechanism. We did in this deep dive by @DirhousssiAmine We instrumented the training loop and decomposed the importance sampling ratio as: log r = α + β, where α is the true policy change (in BF16 space) and β is the precision gap between the training forward pass and a BF16 forward on the same weights. See it like this: α = how much the policy actually changed since the rollout (same precision, different time). β = how much the trainer and inference engine disagree about the same policy (same weights, different precision). The ratio sees α + β and PPO can't tell them apart. Empirically, β is small at the token level (O(1e−2–1e−1)) but it is not an innocent random noise that would wash out over time. We found it to be structured, persistent, and worse for certain tokens: it has a consistent negative bias, correlates with the advantage, and is up to 50x larger on low-probability tokens. However, despite all these concerning properties, none of them explain the mechanism. We saw that just disabling clipping leads to stable convergence meaning that β noise alone does not explain the failure. We tested every plausible explanation and ruled them out one by one: ⭐️ Treating β as pure noise: keeping β but disabling clipping leads to stable convergence. ⭐️ FP32 backward: You're optimizin...

RT Sara Hooker Always happy to collaborate with the @huggingface team 🤗 HF is hitting the milestone of 10M datasets. Now you can easily adapt and evolve your data towards any objective. Shoutout to @ClementDelangue @julien_c @Thom_Wolf @mervenoyann! 🔥

RT Georgia Channing 🏆🏆🏆 EquiformerV3 just topped MatBench Discovery using less than 1/3 of the compute of the closest competitor 🏆🏆🏆 EquiformerV3 precisely simulates chemical physics by scaling SE(3)-equivariant graph attention transformers. AND it was released on @huggingface 🤗 Original tweet: https://x.com/cgeorgiaw/status/2044424403120005601

json is so token inefficient it hurts these days man, these braces and quotes are costing me real $$

favorite AGI/sci-fi vibe these days is coding a robot code together with the robot here vibe-pluging @ElevenLabs in @reachymini for a talk later today

RT clem 🤗 "But here is what we found when we tested: We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models. Those models recovered much of the same analysis. Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens. A 5.1B-active open model recovered the core chain of the 27-year-old OpenBSD bug." https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier Original tweet: https://x.com/ClementDelangue/status/2041953761069793557

RT Julien Chaumond We are giving away Safetensors to the @pytorch foundation (shepherded by the Linux Foundation) Our shared goal is to make the default serialization format for torch safe and performant. To unlock this, governance needs to be independent of @huggingface. Looking forward to more stakeholders contributing to Safetensors in the coming months 🔥 Original tweet: https://x.com/julien_c/status/2041888145587773655

Releasing one of our *largest* robotics project yet in the open We collected and annotated hours of clothes folding with open-arms and collaborators. We then explored how to train the best clothes folding robotic model for bimanual setups. And now we're releasing it all fully in the open: data, code, models, software, explorations, learnings, you name it Enjoy, play with it, use these learnings and share yours! PS: the hub is increasingly *the* place where robotics data is being shared and used, come take a look if you haven't yet. Robotics data has been our fastest growing dataset category by far over the past few months.

RT LeRobot Releasing the Unfolding Robotics blog! Time to unfold robotics: we trained a robot to fold clothes using 8 bimanual setups, 100+ hours of demonstrations, and 5k+ GPU hours. Flashy robot demos are everywhere. But you rarely see the real story: the data, the failures, the engineering. We’re sharing everything: code, data, and details in the blog → https://huggingface.co/spaces/lerobot/robot-folding Original tweet: https://x.com/LeRobotHF/status/2041542790610297259

RT levi Day 93/365 of GPU Programming Studying parallelism today and stumbled upon this incredible blog post/book The Ultra-Scale Playbook: Training LLMs on GPU Clusters by Hugging Face that dives deep into data parallelism, expert parallelism, tensor parallelism, pipeline parallelism and context parallelism. I've read a bit about each of these methodologies before but this is the best resource I've found that really pieces them all together into a unified coherent picture. Kinda like its name implies, the team goes into actual empirical examples based on the 4000 scaling experiments (across up to 512 GPUs!) they conducted. E.g. how does tensor parallelism reduce activation memory for matmuls but still require gathering full activations for LayerNorm? When does pipeline parallelism's bubble overhead outweigh its memory savings? When and why would you combine TP/PP/DP on a specific cluster topology? What's the real memory breakdown between params, gradients, optimizer states and activations and which parallelism strategy targets which? et cetera Also loved all the beautiful and sometimes interactive diagrams that reminded me of http://distill.pub (which makes sense given they used distill's template to create the post). I wish more blog posts in ML would use a similar approach to help visual learners understand the content at an intuitive level. Especially now that rich visualizations/animations are so easy to spin up with LLMs. Really wonderful work by @Nouamanetazi @FerdinandMom @xariusrke @mekkcyber @lvwerra @Thom_Wolf. In times when things are going more and more closed source in, this is such a good example of what great open source AI education and research can look like. Original tweet: https://x.com/levidiamode/status/2041229052804280811

We’re very excited to deepen our work with the @SAIRfoundation co-founded by Terence Tao. We’ve been very active pushing the communities in AI x science in chemistry, physics, biology (more on that very soon) and this aligns perfectly with what the SAIR foundation has been doing in math (and soon extending as well). Sharing datasets, building challenges and communities Exciting future for open-science

RT Lewis Tunstall Terence Tao's SAIR foundation is doing some really cool work on enabling AI4Maths to be open and collaborative I'm heaps excited that we now get to work together on bringing projects like their Mathematics Distillation Challenge to the HF ecosystem. Let's go 🚀! Original tweet: https://x.com/_lewtun/status/2041200203957428659

TFW the R&D boss of arguably the oldest and most legendary robotics lab in the world stops you at a conference to tell you that your robot is "the coolest social robot in the world"

RT Michael Hla I trained an LLM from scratch on pre-1900 text to see if it could come up with quantum mechanics and relativity. While the model is too small to do meaningful reasoning, it has glimpses of intuition. When given observations from past landmark experiments, the model can declare that “light is made up of definite quantities of energy” and even suggest that gravity and acceleration are locally equivalent. I’m releasing the dataset + models and leave this as an open problem to the research community. I also include what this project has taught me about intelligence in a mini essay linked below. 🧵(1/n) Original tweet: https://x.com/hla_michael/status/2039768483018489994

RT Arcee.ai Today we're releasing Trinity-Large-Thinking. Available now on the Arcee API, with open weights on Hugging Face under Apache 2.0. We built it for developers and enterprises that want models they can inspect, post-train, host, distill, and own. Original tweet: https://x.com/arcee_ai/status/2039369121591120030

RT Hynek Kydlíček Oh shit, it seems like all the HF Research team pretraining data has been accidentally leaked to the public. The web, PDFs, and synthetic datasets are expode on hf FineData org... Apparently, an intern used CC to push the data with private=False. Original tweet: https://x.com/HKydlicek/status/2039052059484287299

the LLM is the computer

Who would win when combining best algo(model+optimization)/data of the year? h/t @lvwerra