Ai2 expands collaboration with Hugging Face to accelerate open science
Allen Institute (AI2)
Ai2 just landed a huge storage boost from Hugging Face for its open AI work. That's the backbone behind models already pulled 50M+ times since 2024.
Based on reporting by Allen Institute (AI2) — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ai2 and Hugging Face just deepened a partnership that's less about announcements and more about plumbing. Hugging Face is roughly tripling Ai2's storage on the Hub, pushing it to nearly two petabytes, and lifting the standard rate limits that usually throttle downloads. That means even Ai2's biggest datasets and multi-checkpoint models can move at full speed instead of queuing behind everyone else's traffic.
The scale behind that upgrade is hard to ignore. Since spring 2024, Ai2's models and datasets have been downloaded more than 50 million times from the Hub. And according to Hugging Face's own heatmap, Ai2 publishes more new artifacts each year than any other organization it tracks — a portfolio that now runs past 900 models and 1,200 datasets.
What makes that footprint unusual isn't just volume, it's what's actually in it. Ai2 doesn't stop at releasing model weights. It puts out training data, intermediate checkpoints, data mixes, evaluations, ablations, and demos too, so outside researchers can trace exactly how a model came together rather than just poke at the finished product. Keeping all of that reachable is precisely the kind of bandwidth-heavy problem this expanded deal is meant to solve.
Two examples show how the collaboration plays out in practice. olmOCR-Bench, Ai2's benchmark for document reading — tables, handwriting, multi-column layouts and all — has become Hugging Face's go-to OCR benchmark, wired directly into the Hub's leaderboard so anyone can compare models in one place. Separately, Hugging Face folded Ai2's MolmoAct 2 robotics model into LeRobot this spring, releasing its training data in LeRobot's standard format so it works with low-cost hardware like the SO-100 and SO-101 arms. In just over three weeks, MolmoAct 2's models and datasets were downloaded more than 400,000 times.
The Hub is also becoming the default home for Ai2's science-adjacent work — OlmoEarth for Earth observation, the HiRO-ACE climate model focused on longer-horizon ocean-atmosphere patterns rather than short-term weather, and AstaBench for judging scientific research agents. None of this is flashy in the way a new chatbot launch is, but it's the kind of infrastructure decision that determines whether open research is actually usable by anyone outside the lab that built it.
My take — AI-written commentary, not fact-checked reporting
Plenty of labs slap the word 'open' on a model release and call it a day; Ai2 is doing the unglamorous work of making the entire pipeline — checkpoints, data mixes, evaluations — actually reproducible, which is a much higher bar and a much less marketable one. Tripling storage and killing rate limits sounds boring next to a flashy model drop, but it's the difference between open AI as a slogan and open AI as something researchers can actually build on. The field could use more of this and a lot less chest-thumping about benchmark scores nobody can verify.
Read more about this at: Allen Institute (AI2)