Olmo 3.1 32B Instruct: The Open Recipe, One Generation On

What Olmo 3.1 32B Instruct is

Olmo 3.1 32B Instruct is Ai2's updated instruction-tuned model, Apache 2.0 licensed with a 64K token context window, at roughly $0.20 per million input tokens and $0.60 output through third-party providers. It sits in a family published across two sizes and two variants — 7B and 32B, Instruct and Think — and like everything Ai2 release, its weights, data, training code and evaluations are all public. It is the Instruct counterpart to Olmo 3 32B Think, covered separately here.

Three named post-training stages

Ai2 publish the recipe stage by stage with the dataset behind each one: supervised fine-tuning on Dolci-SFT, direct preference optimization on Dolci-DPO, and a final round of reinforcement learning from verifiable rewards on Dolci-RL, covering mathematics, code, instruction following and general chat queries. RLVR trains against problems whose correctness can be checked mechanically — a maths answer that is right or wrong, a formatting constraint that was or was not met — rather than against a learned reward model guessing at human preference.

LMArena

On LMArena's text leaderboard OLMo 3.1 32B Instruct has 1329 Elo (rank 249); by category, 1382 Elo in Coding (rank 238), 1296 Elo in Math (rank 258), 1327 Elo in Expert (rank 241) and 1293 Elo in Creative Writing (rank 249). Olmo 3 32B Think has 1306 Elo (rank 282).

Context window

The context window is 64K tokens. Each post-training stage is published with its dataset.

Ai2’s results across the stages

Ai2's table shows the effect of each post-training stage on the 32B Instruct model, from SFT to DPO to the final RLVR model: MATH 74.4, 86.6 and 93.4; AIME 2025 8.2, 23.3 and 57.9; LiveCodeBench v3 35.4, 49.6 and 54.7; IFEval 87.7, 87.3 and 88.8; and GPQA 41.3, 47.9 and 48.6. For comparison the table gives OLMo 2 32B at 49.2 on MATH and 36.4 on GPQA. These are Ai2's own figures.

The Olmo 3 family and using it

Olmo 3 models are pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets, and the family runs from the Olmo-3-7B and Olmo-3-32B base models through SFT and DPO checkpoints to the final RLVR models; Ai2 also publish an Olmo 3.1 32B Think. The model needs Transformers 4.57.0 or later and can be loaded in 8-bit with bitsandbytes. Ai2 list a data cutoff of December 2024, and publish fine-tuning recipes through the OLMo-core repository, with DPO and RLVR in open-instruct and evaluation in OLMo-Eval. Because of a checkpointing issue, only the final few RL checkpoints are released for this model. It is intended for research and educational use under Ai2's Responsible Use Guidelines.

Other figures in Ai2’s table

For the final model Ai2 also list AIME 2024 67.8, BigBenchHard 84.0, ZebraLogic 61.7, HumanEvalPlus 86.7, MMLU 80.9, AlpacaEval 2 LC 59.8 and a safety score of 89.5. In the same table Qwen 3 VL 32B Instruct shows MATH 95.1 and AIME 2025 64.2, and Qwen 3 32B without thinking shows MATH 84.3 and AIME 2025 21.3. The model is released under Apache 2.0.

Official Resources