The Allen Institute for AI (Ai2) has released Olmo 3, a series of dense models in 7B and 32B parameter sizes, distinguishing itself through a "model flow" approach rather than just providing open weights. Unlike competitors like Llama or Qwen, which typically only release weights and inference code, Olmo 3 provides the full development lifecycle under an Apache 2.0 license. This includes the Dolma 3 pretraining corpus, the Dolci post-training datasets, and various checkpoints ranging from base and mid-trained to long-context-extended versions. Specifically, the Olmo 3-32B Think variant utilizes inference-time reasoning traces to narrow the performance gap with Qwen 3 32B, notably achieving this while being trained on six times fewer tokens. The release offers three distinct post-training branches: Instruct for chat and tool use, RL Zero for reinforcement learning experimentation starting from the raw base model, and Think for reasoning. Additionally, Ai2 introduced OlmoTrace, a system within the Ai2 Playground that allows researchers to trace model outputs back to specific documents within the training data. This transparency is designed to serve auditors, interpretability researchers, and regulated industries by allowing for direct verification of data contamination and provenance, rather than relying on external inferences.