Allen Institute for AI (Ai2)

A nonprofit institute in Seattle

The Allen Institute for AI, usually shortened to Ai2, is a nonprofit artificial intelligence research institute founded in 2014 by Microsoft co-founder Paul Allen. It is based in Seattle and also has an office in Tel Aviv.^[1] It works as a research organisation rather than as a company selling a product, and a large share of what it builds is released publicly.

Leadership

Oren Etzioni was appointed to lead research in September 2013 and went on to serve as chief executive for nine years, stepping down on 30 September 2022, when Peter Clark became interim CEO.^[1] Ali Farhadi was named CEO on 20 June 2023 and started in the role on 31 July 2023. On 10 March 2026 the institute announced that he was stepping down after about two and a half years, with Peter Clark once again taking over as interim CEO.^[1]

Early projects: search, simulation and a science exam

Among the institute's early projects was Semantic Scholar, an academic search engine that became public in November 2015.^[1] It applied AI to the problem of finding and reading scientific literature, where the volume of published papers makes ordinary keyword search unsatisfactory.

A second early project, AI2-THOR, appeared in 2016 and was described as an embodied-AI simulator: a simulated environment in which software agents learn by acting rather than only by processing text.^[1] The third was Aristo, a system that in 2018 passed an 8th-grade science exam.^[1] Passing a school-level exam is a modest goal by human standards, but it required a system to read a question, reason over background knowledge and pick an answer, which made it a useful public benchmark for how far such reasoning systems had come.

OLMo and the fully open approach

Ai2 announced the OLMo open language model project on 11 May 2023.^[1] On 1 February 2024 it released OLMo 1B and 7B together with the full Dolma training dataset, code, training logs and checkpoints, all under the Apache 2.0 licence.^[1][2] The first batch comprised four final 7B variants that differed in architecture, optimizer and training hardware, plus one 1B model, and all of them were trained on at least 2 trillion tokens.^[6] The institute said its LLM framework was intentionally designed to give access to data, training code, models and evaluation code for open research.^[6]

What Ai2 calls its fully open approach means that training data, code, intermediate checkpoints and evaluations are released alongside the model weights.^[2] Many open models publish only the finished weights: a user can run the model, but cannot see what it was trained on or how it changed during training. Releasing the data, the code and the intermediate checkpoints makes the training process itself available for inspection and reuse.

The model family grows

The line expanded in several directions. OLMoE arrived in September 2024, built on a mixture-of-experts architecture, in which each input is routed to a subset of the model's parameters rather than all of them, and was produced with Contextual AI.^[1] OLMo 2 followed in November 2024, with a 32B version in March 2025. Olmo 3 was released on 20 November 2025, with Olmo 3.1 in December 2025, and Olmo Hybrid on 5 March 2026.^[1][3][5]

Alongside the base models, Ai2 published post-training recipes: Tülu in June 2023, Tülu 2 in November 2023 and Tülu 3 in November 2024.^[1] Post-training is the stage after pretraining in which a model is tuned to follow instructions and behave usefully, so a published recipe describes how that step was carried out. On the multimodal side, Molmo appeared in September 2024 and Molmo 2, aimed at video, in December 2025.^[4]

Where it sits

Ai2's work sits at the meeting point of two traditions: a research institute publishing benchmarks and tools, and an open-source project maintaining a family of released models. Semantic Scholar, AI2-THOR and Aristo belong to the first; OLMo, Tülu and Molmo to the second. The consistent theme is that the artefacts — datasets, simulators, evaluation code, checkpoints — are made public so that other researchers can build on them, which is also what makes the institute's output easy to check against its own claims.