OLMo 3 / Molmo 2 release line

Ai2 released the OLMo 3 family on November 20, 2025: a dense-transformer suite at 7B and 32B parameters with three post-training variants per size, Base, Think (reasoning), and Instruct (chat plus tool use), plus an RL Zero training artifact. Pretraining ran on the Dolma 3 Mix, a 5.9T-token subset of the broader 9.3T-token Dolma 3 open dataset covering web text, scientific literature, and code. All weights ship under Apache 2.0 alongside training data, training code (OLMoCore framework, Open Instruct post-training pipeline), and intermediate checkpoints.

Olmo 3-Think (32B) is reported as the strongest fully open reasoning model in its size class, narrowing the gap to Qwen 3-32B-Thinking while trained on roughly six times fewer tokens. Olmo 3-Instruct (7B) outperforms Qwen 2.5, Gemma 3, and Llama 3.1 on instruction-following benchmarks per the Ai2 release. The supporting toolchain published with the release includes Duplodocus for deduplication, Datamap-rs for data cleaning, OLMES for evaluation, and Decon for test-set decontamination, the same internal pipeline Ai2 uses for its own training runs.

Molmo 2 followed on December 16, 2025: an open multimodal family at 4B, 7B, and 8B parameters, with Molmo 2-O (7B) built on the Olmo backbone so the full vision-language stack is end-to-end open. The 8B model surpasses the prior 72B Molmo in accuracy, temporal understanding, and pixel-level grounding, and the release claims it bests Gemini 3 on video tracking benchmarks. Nine new training datasets totaling more than 9 million multimodal examples (video captions, long-form QA, grounding, tracking, multi-image reasoning) shipped with the model weights.

Among US labs publishing model weights in late 2025, the OLMo and Molmo lineage is the only one that is consistently fully open across data, training recipe, and weights. Most peer releases (Llama, Mistral, Qwen) ship weights without the training corpus, which means independent researchers cannot reproduce the run, audit the data mixture for memorization or contamination, or retrain from scratch with substitutions. Internal Ai2 funding covers the work, so the release line functions as a public-goods publication rather than a grant payable to an external recipient.