AI2 releases OLMo, a fully open language model

The Allen Institute for AI (AI2) released OLMo, a family of four 7-billion-parameter language models each trained on at least two trillion tokens, alongside the complete Dolma pretraining corpus of roughly three trillion tokens, the full training code, and more than 500 intermediate checkpoints per model.

AI2 framed the release as a deliberate contrast with the prevailing use of “open” in the field. Most models described as open-weight, including Meta’s Llama series, published only final model weights and a paper describing methodology in general terms, without the training data itself or the intermediate states that would let outside researchers reproduce or audit how the model had been built. AI2’s blog post put the distinction directly: “open actually means ‘open’” — full insight into training data, code and evaluation methods, rather than the “qualitative assumptions” researchers were left to make about closed or partially open systems.

The practical consequence was reproducibility. With Dolma and the training code public, other researchers could rerun or modify OLMo’s training pipeline rather than only fine-tuning or probing a finished model, and the intermediate checkpoints allowed study of how the model’s capabilities developed over the course of training — data unavailable for essentially any other frontier-scale model at the time, open-weight or closed.

OLMo’s 7B scale put it well behind frontier capability, and AI2 did not claim state-of-the-art performance. Its significance was definitional: it sharpened, rather than settled, an argument already running through 2023 and 2024 about what “open” should mean for a language model, and gave researchers studying training dynamics, data attribution and reproducibility a genuinely inspectable system to work with, in place of the weights-only releases that had come to dominate the term.