AI2 releases Olmo 3 open frontier model family

The Allen Institute for AI released Olmo 3, a family of language models at 7 billion and 32 billion parameters, alongside Base, Think, Instruct and RL-Zero variants at 7B and Base and Think variants at 32B. AI2 described the release as sharing not just final weights but the entire “model flow” — every training stage, intermediate checkpoint, dataset and dependency used to build the models.

That included the underlying Dolma 3 pretraining corpus, of roughly 9.3 trillion tokens, and derived training mixes used for later stages, released with what AI2 said were no license restrictions; the associated code for data processing, training and evaluation; and a tool called OlmoTrace that lets a user trace a model’s output back to the specific training documents that produced it. AI2 described Olmo 3-Base 32B as the strongest fully open base model available and Olmo 3-Think 32B as the strongest fully open “thinking” model, citing wins on benchmarks including SQuAD, DROP and HumanEval against comparably sized open-weight models such as Qwen 2.5, Gemma 3 and Llama 3.1. On mathematical reasoning benchmarks such as AIME, the Think variant traded some raw score for the transparency of a fully documented training and reasoning pipeline.

The release sat apart from most “open-weight” launches, which typically publish final checkpoints without the data or intermediate training artifacts needed to reproduce or audit them. By releasing the full pipeline, AI2 positioned Olmo 3 as evidence that fully reproducible, fully open training at frontier-adjacent scale remained possible even as most well-funded labs, including open-weight-friendly ones, kept their datasets and intermediate checkpoints proprietary.