OLMo 3

The Allen Institute for AI launched OLMo 3 in November 2025 as the most transparent model family in existence. Where most 'open' releases share weights only, Ai2 publishes everything: the Apache-2.0-licensed weights, the Dolma 3 training dataset (9.3 trillion tokens), intermediate checkpoints, training logs, and evaluation pipelines. The family comes in 7B and 32B sizes across four variants: Base, Think, Instruct, and RLZero — with Think being Ai2's first reasoning model, exposing step-by-step thinking traces.

OLMo 3-Think 32B matches Qwen3-32B-Thinking at a sixth of the training tokens, and the 7B Instruct holds its own against models twice its size on instruction following. The 65K context window is modest by 2026 standards — this family optimizes for reproducibility and research, not leaderboard records. For researchers, labs, and anyone who needs to audit what a model was trained on, nothing else comes close.