On November 26, 2024, the Allen Institute for AI (Ai2) released OLMo 2, “a new family of 7B and 13B models trained on up to 5T tokens.” Unlike most open-weight releases that share only the final model, OLMo 2 was fully open: Ai2 said that “because fully open science requires more than just open weights,” it released “weights, data, code, recipes, intermediate checkpoints, and instruction-tuned models.”
Ai2 reported that “OLMo 2 7B and 13B are the best fully-open models to-date, often outperforming open weight models of equivalent size,” and noted that “OLMo 2 7B outperforms LLama-3.1 8B and OLMo 2 13B outperforms Qwen 2.5 7B despite its lower total training FLOPs.” All artifacts, including the technical report, pretraining and post-training datasets, and a demo playground, were made publicly available.
OLMo 2 mattered as a reproducibility benchmark for the field. By publishing the full recipe and data alongside competitive weights, Ai2 let researchers actually study how a strong language model was built, not just use one, providing a transparent counterpoint to the closed and partially open releases that dominate the frontier.