OLMo

OLMo, the Open Language Model from the Allen Institute for AI, is built for reproducible science: alongside model weights, AI2 releases the training code, data recipes, evaluation tools, and intermediate checkpoints at minimum every 1,000 training steps. The models come in 1B, 7B, 13B, and 32B parameter sizes and are trained in two stages, first on trillions of tokens of web data building on the Dolma corpus work, then on smaller curated high-quality mixes, with instruction-tuned variants at each size. OLMo 2's larger models use the newer OLMo-core training framework, and the original repository now points there for current development. Checkpoints are published in both OLMo and Hugging Face formats under the Apache 2.0 license. Researchers reach for OLMo when they need to study how training choices shape model behavior, since the full pipeline is inspectable in a way that closed or weights-only releases are not.