Ai2, the institute releasing fully open language models
Ai2 (the Allen Institute for AI) is a nonprofit research institute based in Seattle. It stands out by releasing not only the weights of its OLMo models, but also their training data, code and intermediate checkpoints.
Data as of October 22, 2026
Key figures
| Country | United States |
|---|---|
| Headquarters | Seattle |
| Founded | 2014 |
| Type | Non-profit |
| Weights policy | Open weights |
| Model families | OLMo, Tülu, Molmo, OLMoE |
| Official site | allenai.org |
A research institute founded by Paul Allen
The Allen Institute for AI, now known as Ai2, was founded in Seattle in 2014 by Microsoft co-founder Paul Allen. It is a nonprofit research institute, not a commercial company. Its mission is “building breakthrough AI to solve the world’s biggest problems.” Its work spans language models as well as AI for science, planetary conservation and robotics.
OLMo, Tülu, Molmo: the model families
- OLMo is Ai2’s family of large language models. As of October 22, 2026, the OLMo 3 generation comes in 7B and 32B sizes, with Base, Instruct (chat) and Think (reasoning-focused) variants. OLMoE is a mixture-of-experts version.
- Tülu 3 is a fully documented post-training recipe combining supervised fine-tuning, DPO and RLVR.
- Molmo is the family of multimodal models that can understand images.
“More than open source”
What sets Ai2 apart is how much it publishes. Many labs release open-weight models without disclosing what data they were trained on. Ai2 releases everything: the training data (the Dolma corpus), the training code (OLMo-core, Open Instruct), intermediate checkpoints and evaluation tools (OLMES). Its OlmoTrace tool even links a model’s answer back to the passages in the training data it may have come from.
In practice, this makes OLMo a reference point for research and teaching: every step can be studied, reproduced or modified. The trade-off is that these models do not try to match the largest commercial models on every task.
How to access the models
- The Ai2 Playground, to chat with the models in your browser.
- Hugging Face for weights and datasets, with OLMo 3 under the Apache 2.0 license.
- GitHub for the code, and partner providers such as Cirrascale for API access.
Learn more at allenai.org. To compare OLMo with other open models, see our page on open-source models.
Frequently asked questions
What is the difference between open weights and fully open?
An open-weight model can be downloaded, but you often don't know what data or methods were used to train it. Ai2 also releases the training data, code, recipes and evaluation tools, so the whole process can be reproduced and studied.
What is Tülu?
Tülu 3 is an open post-training recipe: prompt curation, supervised fine-tuning, preference tuning and reinforcement learning with verifiable rewards. Ai2 publishes the data, code and method, and has applied it to Llama 3 base models among others.
Can I use OLMo in a commercial project?
OLMo 3 is released under the Apache 2.0 license, which allows commercial use. Ai2 notes that its models are intended primarily for research and education, in line with its Responsible Use Guidelines.
How can I try Ai2's models?
The Ai2 Playground lets you chat with the models in your browser. Weights and datasets are on Hugging Face, code is on GitHub, and some partner providers offer API access.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.