Ai2 suits researchers, universities and companies that need to inspect or fine-tune a model's full training recipe rather than just call an API. Its latest release, Olmo-core 3 on October 1, 2026, opens a training stack benchmarked past one trillion parameters, though leadership is interim after March 2026 departures.
The Allen Institute for AI (Ai2) is a Seattle nonprofit research institute founded in 2014 by Microsoft co-founder Paul Allen. It publishes Olmo, a family of language models released with weights, training data, code and intermediate checkpoints, so outside researchers can reproduce and audit how each model was built.
Founded: 2014 · HQ: Seattle, WA, USA · Team: ~200 · CEO: Peter Clark (interim) · Funding: Nonprofit. Funded by the Paul Allen estate and the Fund for Science and Technology; $152M NSF and NVIDIA OMAI award over five years (Aug 2025)
About Ai2
The Allen Institute for AI, known as Ai2, is a Seattle-based 501(c)(3) nonprofit that Microsoft co-founder Paul Allen set up in 2014 to do AI research. Allen appointed Oren Etzioni to direct its research in September 2013, and Etzioni ran the institute for nine years until he stepped down on September 30, 2022. Peter Clark, a leading researcher on the Aristo science-question project, took over as interim CEO, and Ali Farhadi became CEO on July 31, 2023 after leading machine learning at Apple. Farhadi had co-founded Xnor.ai, an Ai2 spinout that Apple acquired in 2020 for an estimated $200 million. Ai2 also keeps an office in Tel Aviv.
Before language models, Ai2 built the Semantic Scholar paper search engine (public since November 2015), the AI2-THOR simulator for embodied agents (2016) and the Aristo system that passed an 8th-grade science exam in 2018. The Olmo line began in February 2024 with 1B and 7B models released alongside the three-trillion-token Dolma corpus, training code, logs and intermediate snapshots. OLMoE, a mixture-of-experts model with about 1B of its 7B parameters active per token, arrived in September 2024. Olmo 2 (7B and 13B) followed in November 2024 and a 32B version in March 2025. Olmo 3 landed in November 2025 as a family of Base, Think, Instruct and RL Zero models, and Olmo 3.1 updated the 32B models from December 2025, including Olmo 3.1 32B Think.
Beyond Olmo, Ai2 publishes the Molmo multimodal models, the Tülu post-training recipes (Tülu 3 in November 2024), OlmoEarth for Earth observation, climate models such as HiRO-ACE, and the Asta suite of research agents with its AstaBench evaluation set. Recent releases include MolmoMotion in June 2026, the 8B-parameter AstaBrief report model and Olmo-core 3 in October 2026. Olmo-core 3 is an open training stack for large mixture-of-experts models; Ai2 says it benchmarked the stack past one trillion total parameters and measured about 2.7 times the throughput of its own earlier MoE implementation in a preliminary 47B-parameter test. On August 6, 2026 Ai2 and Hugging Face expanded their partnership: Hugging Face is roughly tripling Ai2's Hub storage to nearly two petabytes, and Ai2 reports more than 50 million downloads of its models and datasets from the Hub since spring 2024.
Ai2 was first funded by Allen's Vulcan and later by his estate. According to GeekWire, the largest source of money today is the Fund for Science and Technology, a $3.1 billion foundation created under Allen's instructions. In August 2025 the National Science Foundation and NVIDIA committed $152 million over five years ($75 million and $77 million) to the OMAI initiative, which Ai2 leads to build open multimodal models for science. As a nonprofit, Ai2 has no valuation and no equity investors. In March 2026 GeekWire, citing people with knowledge of the situation, reported that the foundation is moving from annual funding to proposal-based support that favors applied AI over building foundation models. A foundation spokesperson said Ai2's work and mission remain the same, and board chair Bill Hilf said all 2026 programs are fully funded.
Peter Clark is interim CEO and Bill Hilf chairs the board, which includes Jody Allen. The team numbers roughly 200 people. Farhadi stepped down in March 2026 to pursue frontier-scale research, and Chief Operating Officer Sophie Lebrecht left with him. Hanna Hajishirzi, a co-lead of the Olmo project and a co-principal investigator on the OMAI grant, and Ranjay Krishna, who led Molmo, were reported on March 23, 2026 by GeekWire to be joining Microsoft's Superintelligence team while keeping their University of Washington faculty roles. In its August 6, 2026 announcement Ai2 still described Clark as interim CEO, and we found no announcement of a permanent successor. Hilf said the board wants a leader who combines scientific depth with nonprofit management experience.
Ai2 frames its mission around AI that tackles large real-world problems, and the first core value it lists is openness: sharing everything it makes. In practice that means publishing the whole model flow, not only weights: pretraining and post-training data, training code (OlmoCore and Open Instruct), evaluation tools (OLMES) and checkpoints. Clark named four focus areas on May 1, 2026: the science of AI system reliability through OMAI, AI for scientific discovery through Asta tools such as ScholarQA and AutoDiscovery, embodied AI through MolmoAct and MolmoBot, and AI for the planet through OlmoEarth, EarthRanger and Skylight.
Ai2's edge is auditability. Open-weight peers such as Qwen3.8 27B from Alibaba Cloud, Gemma 4 31B from Google DeepMind, Phi-4 and Nemotron 3.5 Lightning typically ship weights without the full pretraining corpus, while Ai2 releases data, code and checkpoints, and its OlmoTrace tool traces a model's output back to training data. Its weakness is scale. Hilf told GeekWire that extreme-scale open model research is extraordinarily costly for a nonprofit, and the 2026 departures took several of the people who built the models. Teams that want a vendor-hosted API with enterprise support are better served by commercial labs such as Mistral AI, whose Mistral Large 3 is also open. Browse other research labs on HokAI to compare.
Ai2 says the next generation of Olmo is being built on Olmo-core 3, and in May 2026 Clark said an update on OMAI, the NSF and NVIDIA project to develop the next generation of open models, would come soon. Whether Ai2 keeps pace with larger labs now depends on naming a permanent CEO and on how the Fund for Science and Technology weighs model building against applied science.
Mission
Building breakthrough AI to solve the world's biggest problems.
Products
- Olmo (Fully open language models): https://allenai.org/olmo
- Molmo (Open multimodal models): https://allenai.org/molmo
- Tülu (Open post-training recipes): https://allenai.org/tulu
- OlmoEarth (Earth observation models): https://allenai.org/olmoearth
- Asta (AI research agents and evaluations): https://allenai.org/asta
- Ai2 Playground (Hosted model chat interface): https://playground.allenai.org
Ai2 Models on HokAI
Links
Frequently Asked Questions
How is Ai2 funded?
Ai2 is a nonprofit, so it has no valuation and no venture rounds. Paul Allen's Vulcan funded it first and his estate later; GeekWire reports the biggest backer today is a $3.1 billion foundation named the Fund for Science and Technology. In August 2025 the NSF ($75 million) and NVIDIA ($77 million) committed $152 million over five years to the OMAI initiative that Ai2 leads. GeekWire also reported in March 2026 that the foundation is shifting Ai2 to proposal-based funding, though the board says every planned 2026 program is already paid for.
What does Ai2 release, and what does "fully open" mean?
Ai2 releases Olmo language models, Molmo multimodal models, Tülu post-training recipes, OlmoEarth for Earth observation, and Asta research tools such as ScholarQA and AutoDiscovery. "Fully open" here means Ai2 publishes the pretraining and post-training data, training code and intermediate checkpoints along with the weights. Olmo weights and datasets are hosted on Hugging Face under the allenai organization.
Is Ai2 audited or certified for enterprise security?
As of October 2026 we found no published security audit report, ISO certificate or trust center on allenai.org. Its privacy notice, last updated February 19, 2025, says information is stored and processed in the United States and describes GDPR data-subject rights. Because Olmo weights are downloadable, regulated teams can run the models inside their own audited environment instead of relying on a vendor certification.
Which organizations compete with Ai2 on open models?
Ai2 competes for open-model users with Alibaba's Qwen line, Google's Gemma, Microsoft's Phi and Mistral AI's open releases. Its advantage is publishing training data and checkpoints, which suits research and audit work. Its weakness is scale: Hilf has said extreme-scale open model work is hard inside a nonprofit, so teams chasing peak benchmark scores often look at larger labs.
How do you start using Ai2's models?
Open the Ai2 Playground at playground.allenai.org to chat with an Olmo model in a browser. To build, download weights from huggingface.co/allenai and follow the guides at docs.allenai.org. Ai2 also publishes the Open Instruct post-training pipeline and the OLMES evaluation utility on GitHub for fine-tuning and testing.
Top Alternatives
- NVIDIA: Pick Ai2 for published training data and checkpoints; pick NVIDIA when you need the GPUs and software to train or serve open models.
- Mistral AI: Pick Ai2 for a fully documented model flow; pick Mistral AI for a commercial API with enterprise support.
- Microsoft: Pick Ai2 for nonprofit, fully open research releases; pick Microsoft for small Phi models inside Azure.