Olmo 3 (7B and 32B)
Ai2's Olmo 3 is a fully open 7B and 32B family with Base, Instruct, Think and RL Zero variants, including a 32B reasoning model.
Releases the whole 'model flow', meaning checkpoints, training data and tools for every stage. Think models are trained with RL on verifiable rewards. Ai2 says the base model is 2.5x more efficient to train than Llama 3.1, and context goes up to about 65K tokens.
- Date
- Thursday, 20 November 2025
- Lab
- Ai2
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Training efficiency vs Llama 3.1 | 2.5x Ai2 claim, as relayed by press | company |
Read only via search-result excerpts from these outlets; Ai2's own post and model cards were not opened (fetch quota ran out). Date 2025-11-20 from the press-release URL. Ai2 also shipped Olmo 3.1 later (not dated here).
Sources
- thenewstack.io/ai2s-olmo-3-pushes-the-envelope-for-open-source-llm-performance/
- geekwire.com/2025/ai2-releases-olmo-3-open-models-rivaling-meta-deepseek-and-others-on-per
- infoq.com/news/2025/11/olmo3/
- secure.businesswire.com/news/home/20251120271934/en/Olmo-3-Charting-a-Path-Through-the-Mod
This record was checked against its sources on 6 October 2026. How we check