OLMo 3 releases an inspectable model flow
Ai2 released 7B and 32B OLMo 3 base and reasoning checkpoints with training data, intermediate checkpoints, post-training paths and tools. The release exposes how one base becomes Instruct, Think and RL-Zero variants.
Why it matters: Researchers can intervene at multiple training stages and trace behavior back to data and recipe choices instead of receiving only final weights.