Overview
- Ai2 published Olmo 3 under the Apache 2.0 license with weights, training checkpoints and evaluation assets available on Hugging Face.
- The family includes Base (7B, 32B), Think (7B, 32B) and Instruct (7B), with a reported RLZero variant adapted for reinforcement learning.
- The release exposes end-to-end development details such as data recipes, dependencies and intermediate checkpoints, alongside tools like OlmoTrace for inspecting reasoning traces and linking outputs to training data.
- Curated data drops feature the Dolma 3 pretraining corpus and Dolci post-training suites for SFT, DPO and RLVR, with deduplication and OCR applied to scientific papers.
- Ai2 reports competitive benchmark results for the reasoning-focused Think models and training efficiency gains including 8x SFT throughput and 4x more efficient RL, with performance and efficiency claims awaiting independent validation.