Companies & labs

Ai2 (Allen Institute for AI)

Seattle non-profit AI research institute founded by Paul Allen

8recordsFirst record 2022Latest Oct 2, 2026

Profile

Country
United States
Headquarters
Seattle, Washington, USA
Founded
2014
Founded by
Paul Allen (co-founder of Microsoft)
Type
501(c)(3) non-profit research institute

Ai2 (the Allen Institute for AI) is a non-profit AI research institute in Seattle, Washington, United States. It was founded in 2014 by the late Paul Allen, co-founder of Microsoft, and its first CEO was Oren Etzioni. Under US tax law it is a '501(c)(3)' organisation, run for public-benefit purposes such as science and education. It receives major funding from the Fund for Science and Technology, an Allen entity. After CEO Ali Farhadi stepped down in March 2026, founding member Peter Clark became interim CEO, and Ai2 still introduced him as interim CEO in a post published in May 2026.

Ai2 is not a company that sells products for profit; its aim is 'AI research for the common good'. It publishes training data and code along with its models, so that researchers and developers around the world can examine them and rebuild them. In August 2025 it began a project, supported by the US National Science Foundation (NSF) and NVIDIA, to build fully open AI that helps scientific research. It also runs Semantic Scholar, a service for searching research papers, and Asta, an AI tool that helps with research.

Its best-known work is OLMo, a 'fully open' language model. Ai2 releases not only the model files but also the training data, code and checkpoints saved partway through training. Olmo 3, released in November 2025, comes in two sizes, with 7 billion and 32 billion parameters (the adjustable values inside a model), and is published under the Apache 2.0 licence (a set of rules for use). Molmo, released in 2024, is a model that understands text and images together, and it too uses Apache 2.0. Ai2 says Olmo and Molmo are intended for research and education. Tülu 3 is a model for which Ai2 published the methods and data used to tune AI for conversation; the versions built on Meta's Llama 3.1 follow the Llama 3.1 licence.

Made by AIThis profile was written by AI (Claude) from the 10 sources above (official pages, filings, news reports).How it's made

Records 8

Newest first. Open one to see it with its source.

Allen Institute for AI (Ai2)
Allen Institute for AI releases AstaBrief 8B

The point at which deep-research-style capabilities came down to an 8B open-weight model. Using citation accuracy and recall as evaluation axes matters, because in report generation the problem is not sentence quality bu

AI2
Olmo 3

Released the data and the post-training pipeline together

AI2
Tülu 3

The turning point from preference learning to verifiable rewards

Ai2
Ai2 releases Tülu 3 fully open and formalizes RLVR

Named the approach of replacing the reward model with a verifier function, and released every stage of the pipeline

AI2
Dolma

Released the data together with the tools, making data science reproducible

Akari Asai et al. (University of Washington, AI2, IBM Research)
Self-RAG published

The turning point that moved retrieval from a fixed pipeline to a decision made by the model

UW, AI2 (Wang et al.)
Self-Instruct

A bootstrap in which a model creates its own training data; the starting point of the synthetic data industry

Jaehun Jung et al. (University of Washington, AI2)
Jung et al. present Maieutic Prompting

An early case of treating generated reasoning as something to check and to resolve conflicts in