OLMo 2 / OLMo 3 / OLMo 3.1

Every Olmo 3 and 3.1 checkpoint is downloadable from Hugging Face, including intermediate base, mid-trained, long-context, and post-training checkpoints, plus the data and code artifacts of the model flow.

Weights, code, and tooling ship under Apache 2.0, and Ai2 publishes the training and fine-tuning datasets for download without license restrictions; no separate use policy constrains the grant on the current releases.

Dense 7B and 32B models run from a laptop up to a research cluster with standard open inference stacks and no vendor runtime requirement.

The whole model flow is documented: Dolma and Dolci datasets, the OLMo-core training framework, the open-instruct post-training pipeline, data-cleaning and decontamination tooling, the OLMES evaluation harness, OlmoTrace for tracing outputs back to training data, and stage-level checkpoints for the Instruct, Think, and RL Zero paths.

Strongest reproducibility in this batch: an independent group can rebuild the model from published data, code, recipes, and intermediate checkpoints, and can rerun the published evaluations.

Ai2, a nonprofit institute, develops OLMo in public: OLMo-core, open-instruct, OLMES, datamap-rs, and decon take merged outside pull requests, community calls and Discord shape priorities, and each release explicitly invites the community to validate, critique, and extend the work.