OLMo

OLMo is a language model family from Ai2 for researchers and developers who want to run models on their own hardware or study how they're trained. It provides model weights, datasets, code and reports for examining and adapting the models.

• Chat and tool use: Instruct models handle multi-turn conversations and tool use, with 7B and 32B variants.
• Reasoning: Think models address complex prompts through step-by-step reasoning; 7B-Think exposes intermediate thinking steps.
• Base models: 7B-Base can run on a wider range of hardware, while 32B-Base supports programming, reading comprehension and math tasks, extended contexts and reinforcement learning setups.
• Training data: Downloadable datasets cover pretraining, mid-training and post-training, including web pages, code, books, scientific text and instruction-tuning material where applicable.
• Research tools: OlmoCore supports training, Open Instruct handles post-training, and OLMES provides reproducible evaluation; OlmoTrace links model output to training data.

Downloadable weights support local use, and Ai2 also offers a hosted chat interface on its website. The repository's modeling, training, evaluation and inference code is licensed under Apache 2.0.

The linked Olmo-core repository is Ai2’s current modeling and training codebase; its v3.0.0 release is a software-library version, not an OLMo checkpoint version. The code uses Apache 2.0. Choose the required OLMo generation from Ai2’s model downloads and use that checkpoint’s model card for its version and license.