Olmo 2 32B

OLMo-2 32B is the most capable and largest model in the OLMo 2 family, developed by the Allen Institute for AI (AI2) as part of its bold, research-centric commitment to open AI. Scaling up the same training recipe used for the 7B and 13B versions released in November, OLMo-2 32B boasts 32 billion parameters and is trained on a massive 6 trillion tokens. It’s the first fully open model—with all training data, code, model weights, and documentation publicly released—to outperform GPT-3.5 Turbo and GPT-4o mini across a broad range of multi-skill academic benchmarks. Post-training is executed using Tülu 3.1, combining supervised fine-tuning, direct preference optimization, and reinforcement learning with verifiable rewards (RLVR), yielding an instruction-tuned model that pushes the limits of open-source alignment techniques. Despite its strong performance, OLMo-2 32B is highly efficient: it achieves results comparable to Qwen 2.5 32B while using only one-third the training compute. The model is also supported by OLMo-core, a newly overhauled training framework built for scalability, modularity, and efficiency on modern AI hardware, including support for 4D+ parallelism and asynchronous checkpointing. Every model in the OLMo 2 series—including 7B, 13B, and now 32B—is designed to be finetuned on a single H100 GPU node, making them highly accessible for academic and independent research. With its powerful capabilities, fully transparent design, and optimized training infrastructure, OLMo-2 32B not only sets a new standard for open-weight models but also reaffirms AI2's role as a leading force in democratizing advanced AI.