OLMo
A language model family with downloadable weights, training data and code for people who want to run models locally or study how they're built.
OLMo is a family of language models for researchers and developers who want to inspect, adapt or run a model on their own hardware. Weights are downloadable. Ai2 also provides the training data, code, checkpoints and reports behind the models, giving researchers material to study the full training process rather than only the finished model.
The family includes Base models for further training, Think models that show intermediate reasoning steps, and Instruct models for chat, tool use and multi-turn dialogue. The smaller 7B models can run on a wider range of hardware; the 32B models include options aimed at demanding research and chat tasks. The site also offers a way to chat with OLMo, while downloadable artifacts support work on your own setup.
OLMo's research materials cover pretraining, later training stages and evaluation. The published data includes mixtures of web pages, code, books and scientific text, along with data used for instruction tuning where applicable. Researchers can use OlmoCore for training, Open Instruct for post-training and OLMES for reproducible evaluation. OLMoTrace connects model output to training data. The repository's code is licensed under Apache 2.0.
