OLMo 2 7B: a fully open language model from Ai2

OLMo 2 publishes far more than weights: training code, intermediate checkpoints and the pretraining data mixes are open too. That makes it a strong choice for research into how language models learn.

• Pretrained on roughly 4 trillion tokens in two stages
• 32 layers, hidden size 4096, 4,096-token context
• Intermediate checkpoints let you study training dynamics step by step

Summary written by SolxLab. Weights and code belong to their authors and are distributed under the license shown; always check the official page before use.