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, download, and deploy. According to the announcement, Hugging Face is roughly tripling Ai2's storage on the Hub to nearly two petabytes, beginning immediately. Furthermore, Ai2's traffic will no longer be subject to standard rate limits, allowing large datasets and multi-checkpoint models to download at full throughput. Since the spring of 2024, Ai2's models and datasets have been downloaded over 50 million times from the platform. Hugging Face data indicates that Ai2 publishes more new artifacts annually than any other organization it tracks. Beyond model weights, Ai2 publishes training data, intermediate checkpoints, evaluations, and demos to enable reproducibility. The expanded infrastructure will support ongoing and forthcoming work across multimodal models, tool-using agents, and open science applications. Current projects hosted on the Hub include the Olmo family of language models, Molmo multimodal models, OlmoEarth for Earth observation, and various evaluation tools and benchmarks like olmOCR-Bench and AstaBench. Additionally, Hugging Face integrated Ai2's MolmoAct 2 robotics model into its LeRobot library, helping developers deploy models directly to hardware. Executives from both organizations emphasized that the collaboration strengthens the infrastructure required to keep open-source research artifacts available and easy to reach as Ai2's portfolio continues to grow across multiple domains.