Ai2 releases Olmo 3, with a fully documented 32B Think model under Apache 2.0
Olmo 3 has 7B and 32B Instruct and Think variants. Ai2 releases code, checkpoints and training details alongside the weights.
Ai2 published Olmo 3, a family of 7B and 32B models in Instruct and Think variants. The Hugging Face card here is for Olmo 3 32B Think. The card says long chain-of-thought thinking improves reasoning on math and coding tasks.
The point is the openness. The models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets, and the card lists every stage: base, SFT, DPO and the final RLVR model. Ai2 says it is releasing all code, checkpoints and training details, with logs coming soon. The card requires Transformers 4.57.0 or higher and notes that a newer version, Olmo-3.1-32B-Think, exists.
The licence is Apache 2.0. The card gives no benchmark claims in the parts reviewed, so the quality case rests on the linked material.
The repo creation date, 19 November 2025, is used as the release date. The newer Olmo 3.1 is a separate release.
A model you can retrace from raw data to final weights. That is rarer than a high score.
Sources
Editor's first pass (AI-assisted)
- Real 4.5: Apache 2.0 weights, plus SFT, DPO and base checkpoints named on the card.
- Useful 3.5: Long chain-of-thought for math and code; good for research and for fine-tuning from known stages.
- Notable 4: Open at every stage, with Dolma 3 and Dolci data named; 177 likes only.
- Novel 3: The openness of the pipeline is the point, not the architecture.
Comments