OLMo is the Allen Institute for AI’s fully open large language model project: not just open weights, but the complete modeling, training, evaluation and inference code — plus the training data — needed to reproduce and study the models. The repository has about 6,700 stars on GitHub and was updated in 2025.
Most “open” model releases give you weights and little else. OLMo (Open Language Model) publishes the whole recipe: the ai2-olmo training codebase used by AI2’s own researchers, data pipelines built on the Dolma corpus, evaluation harnesses and inference code. Successive generations — OLMo, OLMo 2 and beyond, in sizes from 1B to 32B parameters — have been released with checkpoints at every stage of training, which makes the repository uniquely valuable for science: you can study how capabilities emerge over training, ablate data decisions, or fine-tune from any intermediate point. The code emphasizes efficient large-scale training with PyTorch, including model-parallel strategies and careful logging, and it is the same stack AI2 uses for its public releases rather than a cleaned-up afterthought. Everything is Apache-2.0 licensed, and the models are free for research and commercial use.
OLMo is completely free and open source under the Apache-2.0 license — models, code and data documentation included. Training at scale requires your own compute.
OLMo is for LLM researchers, students and open-source teams who want to understand and build on a frontier-class model with full transparency into how it was trained.