# Olmo 3 — weights, license and what is actually open

A fully open model flow spanning base, mid-trained, long-context, instruction, reasoning, and RL-zero checkpoints.

Canonical: https://brightaifuture.com/open-models/olmo
Format: model-family
Source publication: Not established
Bright publication: 2026-09-19
Substantive update: None recorded
Evidence and review: AI-assisted primary-source check on 2026-09-19; no independent model reproduction or license certification. Ai2 documents external research using the open stack, including Goodfire’s post-training behavior tracing; those case studies do not validate every capability claim.

## Documented release

Olmo 3 7B and 32B

Organization: Allen Institute for AI (Ai2)

Olmo 3 model-flow release: 2025-11-20 (day precision). Source: https://allenai.org/blog/olmo3

Olmo 3.1 checkpoint update: 2025-12-12 (day precision). Source: https://allenai.org/blog/olmo3

This record begins with the Olmo 3 flow and its cited source also documents the later Olmo 3.1 checkpoint update.

Parameters: 7B and 32B

Architecture: Dense decoder-only transformer with a documented multi-stage training pipeline.

Modalities: text

Context: Up to 65K tokens in the long-context models

## License and commercial use

Apache-2.0 for models and code; data components retain their stated licenses

Model and code licenses permit commercial use; downstream users must review dataset licenses.

## What is open

Weights: available. Final and intermediate checkpoints are released. Source: https://allenai.org/blog/olmo3

Architecture: available. Architecture and changes are documented. Source: https://allenai.org/blog/olmo3

Inference code: available. Model code and deployment instructions are public. Source: https://allenai.org/blog/olmo3

Training code: available. Pre-training and post-training code are released. Source: https://allenai.org/blog/olmo3

Training recipe: available. Training stages, configurations, checkpoints, and mixtures are documented. Source: https://allenai.org/blog/olmo3

Data information: available. Ai2 publishes detailed provenance and composition for Dolma 3 and post-training data. Source: https://allenai.org/blog/olmo3

Training data: available. Dolma 3 and the released post-training datasets are accessible. Source: https://allenai.org/blog/olmo3

Evaluation: available. Evaluation suites, data, and scoring standards are public. Source: https://allenai.org/blog/olmo3

Commercial use: available. Apache-2.0 covers model artifacts; constituent data licenses still apply. Source: https://allenai.org/blog/olmo3

## Uses and strengths described in sources

End-to-end inspectability

Intermediate checkpoints

Open data

Reasoning research

## Hardware and quantization

No universal minimum is asserted; 7B and 32B variants span consumer to multi-GPU deployments.

Community quantizations are available; no single format is treated as canonical here

## Independent evidence

Ai2 documents external research using the open stack, including Goodfire’s post-training behavior tracing; those case studies do not validate every capability claim.

## Limitations

Text-only

Full reproduction still requires substantial compute

## Provenance and history

{}

## Original sources

- [Olmo 3: Charting a path through the model flow](https://allenai.org/blog/olmo3)
- [Olmo model flow](https://allenai.org/olmo)

## Continue exploring

- [Open Models](https://brightaifuture.com/open-models)
- [Open intelligence thread](https://brightaifuture.com/threads/open)
- [Related evidence record](https://brightaifuture.com/discoveries/olmo-2-32b)
- [Related evidence record](https://brightaifuture.com/discoveries/goodfire-olmo-post-training)