Ai2 expands collaboration with Hugging Face to accelerate open science

Ai2 and Hugging Face are expanding their partnership to make fully open AI easier to access and deploy. Hugging Face serves as a central platform for the AI community, and the expanded collaboration gives us substantially more storage and bandwidth to support our growing portfolio of models, datasets, benchmarks, and applications.

Beginning immediately, Hugging Face is roughly tripling our storage on the Hub to nearly two petabytes. Our traffic will also no longer be subject to the Hub’s standard rate limits, allowing even its largest datasets and multi-checkpoint models to download at full throughput. Since spring 2024, our models and datasets have been downloaded more than 50 million times from the Hugging Face Hub.

The added capacity reflects the scale of our presence on the Hub—according to Hugging Face’s official heatmap, we publish more new artifacts annually than any other organization it tracks.

Releasing model weights is only one part of open research. We also publish training data, intermediate checkpoints, data mixes, evaluations, ablations, demos, and other materials that let researchers reproduce our work and examine how our models were built. The expanded infrastructure will help keep those resources available as our portfolio grows, including forthcoming work on multimodal models, tool-using agents, and open models for scientific communities.

"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,” says Clem Delangue, Co-Founder and CEO, Hugging Face. “We are proud to be the infra behind Ai2 to enable their impactful work that varies from MolmoAct VLAs for real robots to OlmoEarth tackling climate and many more: 900+ models, 1,200+ datasets at huge scale."

"Developing models completely in the open is core to how we work at Ai2, and that means publishing everything leading up to a model – as well as the model itself – so the rest of the field can study and build on it," says Peter Clark, Interim CEO, Ai2. "Keeping those artifacts available and easy to reach takes dedicated storage and bandwidth, and this partnership with Hugging Face strengthens our ability to do that as our portfolio of open releases — across many domains — continues to grow."

Hugging Face is already where many people go to find and use our open work. Olmo, our family of fully open language models, lives there, alongside our Molmo multimodal models, OlmoEarth for Earth observation, and AstaBench, our evaluation suite for scientific research agents. The agreement also sets up future open-science work together, building on Hugging Face's efforts to integrate our research artifacts into the Hub and put them in front of wider audiences.

One of the clearest examples of that collaboration is olmOCR-Bench, our benchmark for how well models read real documents: multi-column layouts, tables, handwriting, and captioned charts. Hugging Face's team chose it as their go-to OCR benchmark and wired the dataset into the Hub's leaderboard workflow, so anyone comparing OCR and document-understanding models can see how they score in one place, with direct links to each model.

"olmOCR-Bench helps the community make informed decisions about which OCR model to use, considering model size, whether it's open source, and its license. None of the big household-name LLMs we have today would exist without OCR. olmOCR-Bench has been acting as a north star for their development, so its impact on the field has been huge.” — Nathan Habib, ML Engineer, Evaluation Tooling, Hugging Face

Hugging Face also helps our models run inside the tools developers already use. 

This spring, Hugging Face integrated our open robotics model, MolmoAct 2, into LeRobot – its open-source robotics library – so hobbyists, roboticists, and researchers can fine-tune, evaluate, and deploy the model with familiar workflows and datasets. We also released MolmoAct 2's training data in LeRobot's standard format so developers can take it straight to real hardware, including the low-cost SO-100 and SO-101 arms popular in the LeRobot community. Within just over three weeks of launch, MolmoAct 2’s models and datasets were downloaded more than 400,000 times.

Our evaluation tools live on the Hub too. Benchmarks like IFBench, RewardBench 2, and AstaBench let developers measure how models actually perform at instruction-following, reward modeling, and agentic tasks—reliable measurement that matters more as models multiply and the differences between them get harder to discern, especially for developers choosing which one to build on.

Hugging Face is also becoming a home for our AI-for-science work. The Hub hosts OlmoEarth, our open models for Earth observation, along with climate models like HiRO-ACE—going beyond the more common weather prediction models by focusing on ocean-atmosphere interactions at the long-horizon timescales that matter for the climate. Both are part of Hugging Face's Hugging Science Collabs collection, which tracks artifacts produced through partnerships with the Hugging Science community.

Ai2 and Hugging Face are working toward an ecosystem where anyone can build with, inspect, and conduct experiments on a wide range of AI tools—empowered by the best of Hugging Face’s infrastructure and community, and our uniquely open and scientifically grounded approach to model building. This agreement is a major step in that direction.

Browse our models, datasets, and benchmarks on Hugging Face here.

About Ai2

Ai2 (the Allen Institute for AI) is a Seattle-based nonprofit AI research institute founded in 2014 by the late Paul Allen. Ai2 develops foundational AI research, open models, and tools—including Olmo, Molmo, and OlmoEarth—to advance AI that serves the public good. Learn more at allenai.org.

About Hugging Face

Hugging Face is the platform where the machine learning community collaborates on models, datasets, and applications. Its open-source libraries and the Hugging Face Hub are used by researchers and developers worldwide to build, share, and deploy AI. Learn more at huggingface.co.