Olmo 3 (7B and 32B) and Olmo 3.1

2025-11-20 · Publisher : Ai2 · Domain : Artificial Intelligence

Ai2 releases Olmo 3 (7B and 32B), fully open: data, code, weights and checkpoints. Olmo 3-Think 32B is presented as the best fully open 32B reasoning model.

Ai2 releases Olmo 3 (7B and 32B), fully open: data, code, weights and checkpoints. Olmo 3-Think 32B is presented as the best fully open 32B reasoning model. ### Context After OLMo 2, Ai2 wants to release not only the weights but the entire training path, which it calls the “model flow”. Olmo 3 arrives while comparable open models such as Qwen 3 are released without their data. Olmo 3.1, released on December 12, 2025, extends the reinforcement training of the 32B variants. ### What the model brings - Four variants per size: Base, Think (reasoning), Instruct and RL Zero (7B, for reinforcement research). - Release of the whole flow: the Dolma 3 corpus (about 9.3 trillion tokens in reserve, of which 5.9 trillion for the pretraining mix), Dolma 3 Dolmino (100 billion mid-training tokens), Dolma 3 Longmino (about 50 billion long-context tokens) and the Dolci suite for SFT, DPO and RL. - More economical training: Ai2 states that Olmo 3-Think 32B is trained on about 6 times fewer tokens than Qwen 3 32B for competitive performance, and reinforcement that is 4 times more efficient thanks to in-flight weight updates and continuous batch processing. - OlmoTrace, a tool that links the model's outputs to its training data. - Olmo 3.1 Think 32B: three additional weeks of RLVR, with MATH 96.2, AIME 2024 80.6 and IFEval 93.8 (HF model card). ### Architecture and specifications Dense decoder-only models of 7 and 32 billion parameters, context of about 65,000 tokens. Pretraining on a cluster of up to 1,024 H100s (7.7 thousand tokens per GPU per second for the 7B), mid-training on 128 H100s, post-training on 256 H100s. Data cutoff: December 2024 according to the Think HF model card. License: Apache 2.0 (HF model cards), with responsible use guidelines from Ai2. Full openness: weights, pretraining data, mid-training, long-context and post-training data, code, checkpoints and logs. Technical report: arXiv 2512.13961. ### Cost and access | Item | Detail | |---|---| | Official API | No Ai2 API; demo in the Ai2 Playground | | Hardware | Not disclosed; about 64 GB in BF16 for the 32B (calculated on 32.2 billion parameters); 8-bit quantization possible | | Weights | Base, Think, Instruct, RL Zero on Hugging Face (allenai/Olmo-3-…) | ### Benchmarks | Olmo 3-Think 32B (published by Ai2) | Olmo 3-Think 32B | Qwen 3 32B | Gemma 3 27B | DeepSeek R1 Distill 32B | |---|---|---|---|---| | MATH | 96.1 | 95.4 | 87.4 | 92.6 | | AIME 2024 | 76.8 | 80.8 | 28.9 | 70.3 | | AIME 2025 | 72.5 | 70.9 | 22.9 | 56.3 | | LiveCodeBench v3 | 83.5 | 90.2 | 39.0 | 79.5 | | IFEval | 89.0 | 86.5 | 85.4 | 78.7 | | MMLU | 85.4 | 88.8 | 74.6 | 88.0 | | GPQA | 58.1 | 67.3 | 45.0 | 61.8 | Source: Ai2 blog post of November 20, 2025, Ai2 measurements. A Qwen 3 VL 32B Thinking column exists in the post (MATH 96.7, AIME 2024 86.3). All comparators are open. ### Versus closed models Ai2's table contains no closed model, and no comparison with a closed model was found in the sources consulted. Against Qwen 3 32B, Olmo 3-Think is ahead on MATH (96.1 versus 95.4) and IFEval (89.0 versus 86.5) but behind on LiveCodeBench v3 (83.5 versus 90.2) and on GPQA (58.1 versus 67.3). Ai2 therefore emphasizes transparency and training efficiency rather than a maximum score. For enterprise use, the appeal is full auditability (data traced by OlmoTrace, checkpoints) and the ability to retrain, which closed models and open-weights-only models do not offer. ### Key takeaways Olmo 3 is the reference for fully open models in late 2025: data, code and weights released. Its scores are close to Qwen 3 32B without exceeding it on every axis.

Tags : LLM, Open Source, Entièrement ouvert, Ai2, Olmo 3, Apache 2.0, Raisonnement