Released on November 20, 2025 by Ai2 (the Allen Institute for AI), a US-based non-profit research organization, Olmo 3 is a family of language models. Unlike “open-weight” models that ship only the finished weights, Olmo 3 publishes the training datasets, training code, and intermediate checkpoints as well, all under Apache 2.0. It comes in two sizes — 7B (7 billion parameters) and 32B (32 billion parameters) — with a Base model as the raw material plus derivatives including Think (step-by-step reasoning), Instruct (dialogue and tool use), and RL Zero (a starting point for reinforcement learning). You can try it in the browser through the Ai2 Playground, or download the weights from Hugging Face to run it in your own environment.
Key Features
- “Fully open” — data, code, and checkpoints included: Most open models distribute only the finished weights, but Olmo 3 also publishes Dolma 3, a training corpus of roughly 9.3 trillion tokens, along with the training pipeline and intermediate checkpoints. Outsiders can verify and reproduce how the model was built
- Step-by-step reasoning with the Think models: Olmo 3-Think explicitly generates its reasoning trace before producing an answer. The 32B Think model is regarded as the leading option among fully open models, with particular strength in math, code, and logic problems
- Two sizes: 7B and 32B: The 7B model is manageable on a personal workstation, while the 32B model targets uses that demand higher accuracy. Choose based on your purpose and hardware
- Context length of about 65K tokens: Long-context training (Longmino) is built into the pipeline so performance holds up when you feed in long documents or a sizable codebase at once
- Tool use and multi-turn dialogue: The Instruct variants are tuned for multi-turn exchanges and external tool calls, providing a foundation for agent-style usage
- Trace outputs back to their source with OlmoTrace: In the Ai2 Playground you can trace which parts of a response connect to which parts of the training data — a feature only possible because the training data is public
- Apache 2.0 license: Commercial use, modification, and redistribution are permitted, so you can fine-tune it and build it into your own product
Pricing
| How to use it | Price | What you get |
|---|---|---|
| Ai2 Playground | Free | Try Olmo 3 in the browser, including output tracing via OlmoTrace |
| Downloading the model weights | Free (Apache 2.0) | Get Base / Instruct / Think / RL Zero from Hugging Face and run them yourself |
| Via third-party inference providers | Usage-based | Example: Olmo 3 32B Think on OpenRouter is $0.15 per 1M input tokens and $0.50 per 1M output tokens |
Pricing is current as of August 2026. The models themselves are published free of charge; costs arise from the compute you use to run them yourself, or from a third-party inference provider. Per-provider rates change, so check the official site and each provider’s pricing page for the latest information.
Pros & Cons
✅ Pros
- Because the training data is public, you can verify the basis and biases of the outputs yourself — hard to match for research and education
- Apache 2.0 means you can embed it in commercial services or fine-tune it without negotiating a license
- It can run entirely on your own servers or locally, which suits work where data must not leave your environment
- The Think models surface the reasoning path, not just the answer
- Two sizes (7B and 32B) let you match the model to the GPU capacity you have
⚠️ Cons
- It is a raw model, not a polished service like ChatGPT or Claude; production use requires setting up an inference environment
- In overall capability it does not reach the largest closed models, which are trained with far more compute
- A context length of about 65K tokens is short compared with commercial models that handle a million tokens
- It handles text only — no image generation, no audio
- Japanese-language performance has fewer published evaluations than English, so some use cases will need testing first
Comparison with Similar Services
| Comparison | Olmo 3 | Llama (Meta) | Qwen3 (Alibaba) | Gemma (Google) |
|---|---|---|---|---|
| Provider | Ai2 (US, non-profit) | Meta (US) | Alibaba (China) | Google (US) |
| What is published | Weights + training data + code + checkpoints | Weights only | Weights only | Weights only |
| License | Apache 2.0 | Custom license with conditions | Mostly Apache 2.0 | Custom license with conditions |
| Reasoning-focused model | Think (reasoning traces exposed) | Yes | Yes | Limited |
| Main strength | Transparency, reproducibility, research use | Breadth of ecosystem | Multilingual coverage and a wide size lineup | Lightweight, integrates with Google’s ecosystem |
Who Is It For
- Researchers and university educators who do not want the model to remain a black box
- Developers in regulated industries who must explain model behavior back to the training data and procedure
- Engineers who want their own language model, fine-tuned on their own data
- Teams that cannot send confidential information to an external API and need inference that stays in-house
- Learners who want to study how a large language model is actually built, from real data and code
Summary
Olmo 3 pushes back on a situation where “open source AI” had come to mean little more than distributing weights, by publishing the entire process — data, code, and checkpoints. It is not a ready-made chat service, so ChatGPT or Claude will be easier for general use, but when transparency, reproducibility, or self-hosting are requirements, Olmo 3 becomes a strong option. Start by checking the response quality of the 32B Think model in the Ai2 Playground, and if it fits your purpose, pull the weights from Hugging Face and run them locally.