OLMo 2 13B

OLMo 2 13B is the larger of the two models in Ai2's OLMo 2 family, released in November 2024. Trained on 5 trillion tokens of curated public data, it competes with leading open-weight models like Llama 3.1 8B on standard benchmarks while meeting the Open Source Initiative's definition of open source AI.

What sets OLMo 2 apart is the completeness of its release: model weights, pre-training data (Dolma), training code, intermediate checkpoints, and transparent evaluations are all publicly available. Post-trained variants (SFT, DPO, and RLVR instruction versions) were released under Apache 2.0, making this one of the most reproducible frontier-adjacent models in existence.