Overview
- Ai2 published model weights alongside pretrained, mid-training and post-training checkpoints under Apache 2.0, with downloads available on Hugging Face.
- The family includes Base (7B, 32B), Think (7B, 32B), Instruct (7B) and RLZero variants targeting core skills, step‑by‑step reasoning, chat/tool use and reinforcement‑learning research.
- Olmo 3‑Think produces inspectable reasoning traces and, according to Ai2’s reporting, reaches competitive results with roughly one‑sixth the training tokens of peers like Qwen 3.
- Data releases feature the Dolma 3 corpus—9.3 trillion tokens overall with a 5.9 trillion‑token pretraining mix—plus documented curation, decontamination and evaluation recipes.
- Ai2 also released tooling for traceability and reproducible research (OlmoTrace, Olmo‑core, datamap‑rs, OLMES), as Nvidia publicly praised the effort, which lands as U.S. teams debate rebuilding from scratch versus forking Chinese base models like DeepSeek.