The Allen Institute for AI (AI2) has announced Olmo 3, a new family of open-source language models, followed by an update to Olmo 3.1. The key innovation is the release of the complete "model flow"—the entire development pipeline—not just the final models.

The Olmo 3 Model Family

Instead of one model, AI2 provides specialized models built from a common base. All components (data, code, weights) are fully public.

1. Olmo 3-Base (7B, 32B): A powerful foundation model for further customization.
2. Olmo 3-Think (7B, 32B): Specialized for complex, step-by-step reasoning and problem-solving.
3. Olmo 3-Instruct (7B): Optimized for chat, following instructions, and using tools.
4. Olmo 3-RL Zero (7B): Designed as a starting point for reinforcement learning research.

The Olmo 3.1 Update

This update significantly improves the 32B models:

1. Olmo 3.1 Think 32B: An enhanced reasoning model showing major gains on math and logic benchmarks.
2. Olmo 3.1 Instruct 32B: A new, larger instruction-tuned model for chat and dialogue, noted as AI2's strongest open 32B chat model.

Performance Claims

AI2's evaluations position Olmo 3 as a top contender in the open-source space:

1. Olmo 3-Base 32B is called the strongest fully open base model.
2. Olmo 3-Think 32B is noted as the strongest fully open thinking model, competing closely with leading open-weight models like Qwen 3.
3. Olmo 3-Instruct 7B is said to match or outperform comparable models like Qwen 2.5 and Gemma 3.

The "Model Flow" Concept

This is the core idea behind the release. AI2 is providing not just model snapshots, but the entire developmental journey—including intermediate checkpoints and datasets. This aims to allow researchers and developers to:

Understand how capabilities develop.

Fork and customize the model at any stage.

Achieve full reproducibility and transparency.

Availability

You can try the models in the AI2 Playground or download the weights from Hugging Face. API access is planned for the future.

In short, with Olmo 3 and 3.1, AI2 is advancing open-source AI by combining competitive model performance with unprecedented transparency into the entire creation process.

If you're interested, I can provide more details on specific model capabilities or benchmarks mentioned in the report.