Olmo 3: AI2’s Fully Open Language Model Flow

The Allen Institute for AI (AI2) has released Olmo 3, a fully open-source large language model ecosystem designed for transparency, accessibility, and research. Unlike typical model releases, Olmo 3 shares the entire development flow — from datasets and checkpoints to training tools — allowing the community to inspect, adapt, and extend the models at every stage.

Olmo 3 includes several variants:

• Base (7B & 32B): strong foundational models, top among fully open LLMs.

• Instruct: tuned for chat and instruction.

• Think: optimized for reasoning and intermediate steps.

• RL Zero: supports open reinforcement learning training.

The models are trained on large-scale, documented datasets like Dolma 3 and refined with specialized mixes for reasoning and long-context tasks. Tools like OlmoTrace offer traceability from model output back to training data.

Olmo 3 aims to democratize high-performance AI while promoting reproducibility, inspection, and ethical development.

Source: Olmo 3 blog post.

Image:  Sora