Ai2 releases Olmo 3 with open reasoning models

Background

Ai2, the Allen Institute for AI, builds language models and publishes them with their construction details, not only the final weights. Olmo is the name of its model family. Ai2 uses the phrase fully open for models whose development process is shared along with the model itself, and Olmo 3 was the release that carried that approach into a new generation.

What happened

On 20 November 2025 Ai2 released the Olmo 3 family in two sizes, 7B and 32B parameters^[1]. Parameters are the numerical values a model adjusts during training, and the parameter count gives a rough sense of model size. The family includes four variants: Base, Think, Instruct and RL Zero^[1]. Each is a separate model in the same family, so a reader can compare how the variants differ in use.

Key details

Pretraining used a new dataset called Dolma 3, of about 9.3 trillion tokens^[1]. A token is a small unit of text, often a whole word or part of one, that a language model reads and predicts. Pretraining is the stage in which a model learns general patterns from a large body of text before any later tuning.

Ai2 described Olmo 3-Think 32B as the strongest fully open thinking model^[1]. A thinking model is one that works through intermediate reasoning steps before it gives a final answer.

The release also introduced OlmoTrace, a tool that lets users trace model outputs back to the training data^[1].

Follow-up

An update, Olmo 3.1, followed on 12 December 2025^[1]. The update kept the same family name.

Impact

Olmo 3 showed that a lab can publish a family of reasoning-capable models at two sizes together with a dataset, a trace tool and a set of variants. For students and researchers, the release is useful because the parts of the system can be studied and reused, not only the final chat behaviour. Later Ai2 releases, including Molmo 2 and Olmo Hybrid, were built on the Olmo family.

Limits

Phrases such as strongest fully open thinking model are the developer's judgement and should be checked against later independent evaluations before they are relied on.