Hugging Face triples storage for Ai2 open-source models

The Allen Institute for AI (Ai2) has expanded its partnership with Hugging Face to significantly scale up the hosting of its open-science assets. Under the new agreement, Hugging Face is tripling Ai2's storage capacity on its Hub to nearly two petabytes. Crucially for developers, Ai2's traffic will bypass standard rate limits, allowing high-throughput downloads of massive datasets and multi-checkpoint models. This infrastructure upgrade supports a massive portfolio that already includes more than 900 models and 1,200 datasets, which have garnered over 50 million downloads since spring 2024.

The expanded capacity will house Ai2's diverse releases, including the Olmo language models, Molmo multimodal models, and OlmoEarth for Earth observation. It also supports evaluation suites like AstaBench, IFBench, and RewardBench 2. A key highlight of the collaboration is olmOCR-Bench, a document-analysis benchmark that Hugging Face has integrated directly into its leaderboard workflow. Additionally, the partnership supports AI-for-science initiatives like HiRO-ACE, a climate model tracking ocean-atmosphere interactions, which is hosted alongside OlmoEarth in the Hugging Science Collabs collection.

For practitioners, this collaboration translates to frictionless deployment and deeper research capabilities. Developers can access intermediate checkpoints and training data to fully reproduce Ai2's work. In robotics, Ai2 integrated its MolmoAct 2 model into Hugging Face's LeRobot library, formatting training data for low-cost hardware like the SO-100 and SO-101 robotic arms. This integration led to over 400,000 downloads of MolmoAct 2 models and datasets within three weeks of launch. By removing bandwidth bottlenecks and integrating benchmarks directly into the Hub, the partnership allows engineers to easily evaluate, fine-tune, and deploy highly complex models.