Ai2 and Hugging Face announced an expansion of their collaboration aimed at making fully open artificial intelligence more accessible and reproducible. Why should you care? Because it’s not just about uploading model weights: it’s about keeping training data, intermediate checkpoints, evaluations and tools available so you can replicate and understand how models are built.
What they announced
-
Hugging Face triples Ai2’s storage on the Hub to nearly 2 petabytes, and removes rate limits that made large downloads painful.
-
Since spring 2024, Ai2’s models and datasets have been downloaded more than 50 million times from the Hub. Ai2 now publishes more new artifacts each year than any other organization tracked by Hugging Face.
-
The agreement supports 900+ models and 1,200+ datasets from Ai2, including families like Olmo (language models), Molmo (multimodal) and OlmoEarth (Earth observation).
-
Concrete integrations: MolmoAct 2 (a robotic control model) was integrated into LeRobot for fine-tuning and deployment on real hardware; MolmoAct 2 saw over 400,000 downloads in a few weeks.
Why this accelerates open science
Publishing weights isn’t enough to reproduce results. Ai2 releases training scripts, intermediate data, checkpoints and evaluations. With more storage and bandwidth those artifacts stay accessible as projects grow in size and complexity.
-
Unlimited downloads are crucial for multi-checkpoint models and huge datasets: they let you grab the full training history to reproduce ablations or retrainings.
-
Centralized leaderboards and benchmarks make side-by-side comparisons simple. One immediate example is olmOCR-Bench, which measures reading of real documents (columns, tables, handwriting, tables with captions). By integrating it into the Hub, any researcher can compare models in one view and access both the artifact and the evaluation code.
olmOCR-Bench helps the community decide which OCR model to use considering size, license and openness. It’s been a key reference in the improvement of many models.
Nathan Habib, ML Engineer, Evaluation Tooling, Hugging Face
Technical impact for developers and researchers
-
Infrastructure: almost 2 PB and throughput without throttling reduce friction when downloading big checkpoints and massive datasets needed for replication experiments and fine-tuning.
-
Workflow: integrating models into libraries like LeRobot makes it easier to take a model from the Hub to real hardware. For MolmoAct 2 that meant being able to test on inexpensive arms like the SO-100 and SO-101 with standardized data formats.
-
Evaluation and metrics: benchmarks like IFBench, RewardBench 2 and AstaBench let you measure instruction following, reward modeling and agentic tasks. Those metrics matter as differences between models become more subtle.
-
Science and climate: OlmoEarth and HiRO-ACE appear in the Hugging Science Collabs collection, making it easier for climate and ocean researchers to access models and data for long-horizon experiments.
Risks and open questions
-
Centralization of infra: relying on a single platform for storage and delivery is practical but creates operational and governance dependencies. What happens if policies or pricing change?
-
Costs and sustainability: keeping petabytes of data and fast downloads isn’t free. The current collaboration mitigates this for Ai2, but there’s a bigger question about funding models for open science at scale.
-
Reproducibility vs compute capacity: publishing checkpoints is valuable, but reproducing large experiments requires training capacity that not everyone has. Still, open artifacts enable inspection and partial evaluations that improve transparency.
Concrete actions you can take now
-
If you’re a developer: explore Olmo and Molmo on Hugging Face, try deployments and fine-tuning with LeRobot integrations or Hub pipelines.
-
If you’re a researcher: use benchmarks like AstaBench or olmOCR-Bench to evaluate models in a comparable and reproducible way.
-
If you’re an instructor or student: the open checkpoints and datasets are great material for ML courses and final projects that need real replication.
-
If you manage infrastructure or policy: think about how to balance open access with financial sustainability and multilateral governance.
Ai2 and Hugging Face aren’t just enlarging a repository: they’re expanding the community’s capacity to inspect, measure and build on open work at scale. That doesn’t guarantee everything in the field will be perfect, but it does lower barriers so more people can research in a verifiable and responsible way.