Ai2's Olmo 3 ships a fully open 32B thinking model, with the data and checkpoints attached
Ai2 released Olmo 3 on 20 November as a family of dense 7B and 32B models under Apache 2.0, with a 65,000-token context window — 16x Olmo 2's — and weights on Hugging Face alongside a hosted demo. The stack is Dolma 3, a roughly 9.3-trillion-token pretraining corpus, plus the Dolci post-training datasets, and Ai2 describes Olmo 3-Think 32B as the strongest fully open thinking model at that scale. Olmo 3-Base 32B outperforms other fully open base models, including Stanford's Marin and Apertus.
The headline is not the parameter count, it is the license plus the completeness: Apache 2.0 means no gating, no monthly-active-user threshold and no side letters, so the weights can ship inside a commercial product. Ai2 also published every intermediate training checkpoint, the training code and the pretraining corpus, which is what turns reasoning traces into something traceable back to the data and the training decisions that produced them. At roughly 20GB in 4-bit, a 32B dense model fits a single 24GB GPU.
WHY IT MATTERS
This is a model you can legally self-host, fine-tune and audit end to end, so no vendor can deprecate it, re-price it or rate-limit it under you. The released checkpoint series make it a practical base for long-horizon reasoning experiments without a frontier lab's budget, and full data provenance is the thing regulated buyers and eval teams ask for before a model touches production. Be clear-eyed about the ceiling: 32B dense narrows the gap to the best open-weight models rather than closing it — Kimi K2 Thinking is still the open benchmark leader.
