Olmo 3 32B Think: Ai2's Fully Documented Reasoning Model

What Olmo 3 32B Think is

Olmo 3 is Ai2's model family, released in 7B and 32B sizes with both Instruct and Think variants. This is the largest reasoning variant, Apache 2.0 licensed and self-hostable, with a 64K token context window and pricing around $0.15 per million input tokens and $0.50 output through third-party providers. Ai2 describe the Think variants' long chain-of-thought as improving reasoning tasks like mathematics and coding.

The post-training is named, dataset by dataset

The Think variants were produced by supervised fine-tuning on the Dolci-Think-SFT dataset followed by direct preference optimization on Dolci-Think-DPO, both covering mathematics, code, chat and general knowledge queries, both published by name. Ai2 release the weights, data, training code and evaluation.

LMArena

On LMArena's text leaderboard OLMo 3 32B Think has 1306 Elo (rank 282); by category, 1360 Elo in Coding (rank 261), 1312 Elo in Math (rank 241), 1303 Elo in Expert (rank 261) and 1264 Elo in Creative Writing (rank 285).

Context window

The context window is 64K tokens. Long chains of thought consume the same budget as the input.

Ai2’s results

For Olmo 3 32B Think, Ai2 report MATH 96.1, AIME 2024 76.8, AIME 2025 72.5, LiveCodeBench v3 83.5, HumanEvalPlus 91.4, IFEval 89.0, MMLU 85.4 and GPQA 58.1. The same table lists Qwen 3 32B at 95.4, 80.8, 70.9, 90.2, 91.2, 86.5, 88.8 and 67.3, and DeepSeek-R1-Distill-Qwen-32B at 92.6 on MATH and 56.3 on AIME 2025. Across the SFT, DPO and final RLVR stages AIME 2025 moves from 66.2 to 70.7 to 72.5. These are Ai2's own evaluations.

Using it

Ai2 evaluated with, and recommend for generation, a temperature of 0.6, top_p of 0.95 and a maximum of 32,768 tokens. The model needs Transformers 4.57.0 or later, can be served with vLLM, and can be loaded in 8-bit with bitsandbytes. Intermediate post-training checkpoints are published as revisions named step_XXX. The final Think models were trained with reinforcement learning from verifiable rewards on the Dolci-Think-RL dataset of mathematics, code, instruction-following and chat queries, after supervised fine-tuning and DPO. Ai2 list a data cutoff of December 2024, and note that the models have no safety filtering, can be prompted to produce harmful content, and are intended for research and educational use under Ai2's Responsible Use Guidelines. Training code is in OLMo-core and open-instruct, and evaluation code in OLMo-Eval.

The Olmo 3 release

Olmo 3 is a family of 7B and 32B models in Instruct and Think variants, pre-trained on the Dolma 3 dataset and post-trained on the Dolci datasets. The Think model is built from the Olmo-3-32B base through Olmo-3-32B-Think-SFT and Olmo-3-32B-Think-DPO checkpoints, each published separately, and Ai2 later released Olmo 3.1 32B Think. Ai2 also list IFBench 47.6, ZebraLogic 76.0, AlpacaEval 2 LC 74.2 and a safety score of 68.8 for the final model. It is released under Apache 2.0.

Licence and contact

The model is developed by the Allen Institute for AI and described as a Transformer-style autoregressive language model. Ai2 note that its statements are often inaccurate and facts should be verified. Ai2 list olmo@allenai.org for technical enquiries and publish the Olmo 3 paper on their site.

Official Resources