AI2 announces OLMo 3: first fully open thinking model family from 7B to 32B

AI2 (Allen Institute for AI)/OLMo

AI2 released OLMo 3, a family of fully open-weight language models trained on Dolma 3 corpus, with both 7B and 32B parameter variants in base, thinking and instruct configurations. The 32B thinking model is the first fully open reasoning model, generating explicit step-by-step reasoning chains. OLMo 3 provides complete transparency: weights, training code, datasets and intermediate checkpoints under Apache 2.0.

• OLMo 3 32B Think model trained on ~6 trillion tokens with 65,536 token context window
• Full model flow released: weights, training code, datasets, and training pipeline documentation
• Dolma 3 training corpus includes curated scientific and technical content for improved reasoning
• Thinking variants generate explicit reasoning chains for complex logic and instruction-following