Olmo 3
lastmod 2025-11-24created 2025-11-241 edits
A new path in the model flow for leading open-source AI
About
- Olmo 3 releases not only the model's final result but the entire development process (model flow), providing full traceability down to data, code, and checkpoints
- Consists of four models at 7B and 32B parameter scale, Base, Think, Instruct, and RL Zero, usable for various research purposes such as reasoning, chat, and reinforcement learning
- Based on the Dolma 3 and Dolci datasets, releases transparent training data of about 9.3 trillion tokens covering web, code, math, science, and more
- The OlmoTrace tool can trace in real time which training data a model's output came from, strengthening transparency and trustworthiness
- Fully open-source release lets anyone intervene, modify, and retrain at specific stages of the model, building a verifiable AI research ecosystem