OLMo-3-7B

OLMo-3-7B is a 7B-parameter language model from the OLMo family (Allen Institute for AI, Ai2). It is fully open: weights, training data (Dolma 3), code (the OLMoCore framework and Open Instruct pipeline), and complete training recipes are all publicly available, ensuring full reproducibility.

It comes in several variants: 7B-Base (a foundation model optimized for efficient inference), 7B-Think (reasoning-focused, surfacing intermediate thinking steps), 7B-Instruct (tuned for chat, tool use, and multi-turn dialogue), and an RL Zero track. The model delivers strong reasoning while remaining highly efficient at 7B scale.

The full training pipeline covers pretraining on Dolma 3 data, midtraining, and post-training (supervised fine-tuning, DPO, and reinforcement learning). The weights are released under the Apache 2.0 license.