OLMo 3 32B Think is a large-scale reasoning model developed by the Allen Institute for AI (Ai2) as part of its third-generation Open Language Model (OLMo) family. As a specialized variant optimized for logic and problem-solving, this 32-billion parameter model utilizes an explicit chain-of-thought mechanism to decompose complex tasks. It is built as a "glass-box" model, emphasizing full transparency by providing access to its training data, code, and post-training recipes.
The model's architecture is a decoder-only transformer featuring 64 layers and grouped-query attention to optimize inference. It was pre-trained on the Dolma 3 corpus, which consists of approximately 6 trillion tokens heavily weighted toward mathematics, scientific literature, and source code. Following pre-training, the Think variant underwent reinforcement learning from verifiable rewards (RLVR) on the Dolci-Think-RL dataset to enhance its ability to generate multi-step reasoning traces.
Designed for long-context tasks, OLMo 3 32B Think supports a context window of 65,536 tokens. It achieves high performance on reasoning-intensive benchmarks such as MATH and HumanEval+, frequently surpassing larger open-weight models of its class. Along with the model weights, Ai2 provides provenance tools like OLMoTrace, which allows researchers to trace model responses back to specific segments of the original training corpus.