OLMo 3

7B and 32B dense Transformers with 65K native context (up from 4K). 3-stage training: main pretraining (~95%), mid-training (100-200B tokens of code/math/QA/thinking), and long-context extension. Trained on Dolma 3 (~9.3T token pool including olmOCR-processed science PDFs). 5.5-5.9T tokens used.

32B Think: MATH 96.1, AIME 2024 76.8, AIME 2025 72.5, HumanEvalPlus 91.4, MMLU 85.4. OLMo 3.1 (Dec 2025) extends RL training (+3 weeks RLVR): AIME 2025 78.1, IFEval 93.8. AA Intelligence: 14 (3.1 Think). Strongest fully open thinking model at release. Apache 2.0.