OLMo

OLMo, the Open Language Model from the Allen Institute for AI, was released in February 2024 as an answer to a specific frustration: most 'open' models ship weights while withholding training data, code and logs. Ai2 went the other way, publishing the 1B and 7B model weights alongside the 3-trillion-token Dolma training corpus, training and evaluation code, intermediate checkpoints, and training logs. Project lead Hanna Hajishirzi compared studying LLMs without training data to drug discovery without clinical trials.

Everything is under Apache 2.0, permitting commercial use, and OLMo 2 later met the Open Source Initiative's definition of open-source AI, a bar most open-weight releases miss. The 7B and 13B OLMo 2 models were trained on 5 trillion tokens and reported as competitive with Llama 3.1 8B. For researchers, OLMo is a scientific instrument: the checkpoints and logs let anyone study how capabilities form during training. For builders, it is one of the few frontier-adjacent models with genuinely clean commercial licensing. openWeights is true, and unusually, so is everything else.