OLMo is a family of fully open language models from Ai2 (the Allen Institute for AI), released with its training data, code and checkpoints, not just its weights.
In plain words
Most open models share only the finished weights, so nobody outside can see exactly how they were made. OLMo shares the whole recipe: the pre-training data, training and evaluation code, logs and many in-between checkpoints. That lets researchers study how a model's abilities and biases develop. The first OLMo came out in February 2024, OLMo 2 added 7B, 13B and 32B sizes, and Olmo 3 includes base, reasoning and chat variants.
A quick example
A researcher wondering why a model repeats a certain fact could search OLMo's actual training data for that fact and compare checkpoints from different points in training, which is impossible with weights-only models.
Where you'll meet it
- Research on how language models learn.
- A fully transparent base for fine-tuning.
Related
Sources
- PaperOLMo: Accelerating the Science of Language Models, Groeneveld et al. (arXiv) · read 29 Sept 2026
- Paper2 OLMo 2 Furious, Team OLMo (arXiv) · read 29 Sept 2026
- DocsOlmo, Ai2 · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.