Ai2 and Hugging Face Expand Partnership, Unlocking 2 Petabytes for Open AI

The Allen Institute for AI (Ai2) and Hugging Face have expanded their partnership well beyond a storage upgrade. Ai2 is the most prolific publisher of open AI artifacts on the Hub, and Hugging Face is backing that output with infrastructure to match.

What changed, concretely

The terms are effective immediately. Hugging Face is roughly tripling Ai2's storage on the Hub to nearly two petabytes, and Ai2's traffic will no longer be subject to the Hub's standard rate limits. That means even its largest datasets and multi-checkpoint models can download at full throughput. Rate limiting has historically been a real friction point for teams trying to pull down large model checkpoints or multi-hundred-gigabyte datasets, particularly for groups without major cloud infrastructure behind them.

Since spring 2024, Ai2's models and datasets have been downloaded more than 50 million times from the Hub. That number explains why the upgrade was necessary: Ai2 was outgrowing the default tier.

The most prolific publisher on the Hub

According to Hugging Face's own model release heatmap, Ai2 publishes more new artifacts annually than any other organization it tracks, ahead of commercial labs with far larger headcounts and compute budgets. Ai2's Hugging Face profile currently lists 968 models and 1,275 datasets.

Those numbers only tell part of the story. Releasing model weights is one piece of open research. Ai2 also publishes training data, intermediate checkpoints, data mixes, evaluations, ablations, and demos, giving researchers everything they need to reproduce results and inspect how models were built.

Why the "fully open" distinction matters

Ai2 sits in a narrow category: organizations capable of training frontier-scale foundation models that are neither commercial labs nor national security programs. Most open-weight releases from commercial labs, including Meta's Llama and Mistral's models, stop at weights. Ai2 publishes the full stack.

Hugging Face CEO Clem Delangue described it directly: "Ai2 has been showing us what 'fully open' looks like, sharing everything end-to-end from training scripts to intermediate datasets behind impactful models like OLMo and Molmo." That endorsement carries weight coming from the operator of the dominant model distribution platform in the field.

The robotics model worth watching

The clearest example of deeper integration is MolmoAct 2, Ai2's open robotics foundation model. Its inference and training pipelines are built directly into Hugging Face's LeRobot platform, so teams already working in that ecosystem can drop the model into their existing setup without retooling. The model was downloaded more than 400,000 times within weeks of release.

MolmoAct 2 also released 720 hours of bimanual manipulation data, code, and evaluations in full, a contrast to competitors like NVIDIA's GR00T and Physical Intelligence's π0.5, which keep their training data closed. The performance gap from the first generation is significant: inference latency dropped from 6,700 milliseconds per action to 180, a 37-fold speedup that crosses the threshold required for real-time robot control.

Benchmarks as shared infrastructure

Hugging Face has integrated Ai2's evaluation tools into the Hub's leaderboard system. olmOCR-Bench, Ai2's benchmark for how well models read real documents including multi-column layouts, tables, handwriting, and captioned charts, is now the Hub's default OCR benchmark, wired into the leaderboard workflow so anyone comparing document-understanding models sees scores in one place.

Other Ai2 evaluation tools available on the Hub include:

• RewardBench 2 — measures how well reward models rank outputs, critical for RLHF pipelines
• IFBench — evaluates instruction-following behavior
• AstaBench — tests performance on agentic scientific research tasks

As model counts multiply and quality differences narrow, reliable benchmarks are the infrastructure that lets builders make informed decisions about which model to deploy.

Open science at petabyte scale

The expanded storage and bandwidth will matter as Ai2's science-focused model families grow. The Hub already hosts OlmoEarth, Ai2's open models for Earth observation, and climate models like HiRO-ACE, which targets ocean-atmosphere interactions at the long-horizon timescales relevant to climate research rather than near-term weather prediction. Ai2 has also brought online a new computing system funded by Nvidia and the National Science Foundation, the first milestone in a $152 million project to build open AI models for scientific fields including materials science, biology, and energy.

Forthcoming work flagged in Ai2's announcement includes multimodal models, tool-using agents, and open models for scientific communities, all of which will land on the Hub first. For anyone building on open models, Ai2's Hugging Face page is the place to watch.