Allen Institute for AI
Last reviewed
Sources
46 citations
Review status
Source-backed
Revision
v3 · 7,814 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
46 citations
Review status
Source-backed
Revision
v3 · 7,814 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Allen Institute for AI (branded since the early 2020s as Ai2, and historically abbreviated AI2) is a Seattle-based non-profit artificial intelligence research institute, founded in 2014 by the late Microsoft co-founder Paul Allen, that builds fully open foundation models and releases not just model weights but the entire training pipeline behind them.[1][2] Its flagship OLMo language-model series is widely described as the most thoroughly open frontier model family ever released, because Ai2 publishes the weights, the pretraining data, the training code, every intermediate checkpoint, the training logs, and the evaluation harness under permissive licenses.[12][13] Paul Allen seeded the institute with an initial commitment of more than $125 million, and it now employs roughly 200 to 250 researchers, engineers, and staff from a headquarters in Seattle's Eastlake neighborhood with a satellite office in Tel Aviv, Israel.[1][2]
Ai2 distinguishes itself from peer laboratories by releasing not only model weights but also pretraining data, training code, intermediate checkpoints, training logs, and evaluation infrastructure.[12][13] That commitment to radical openness, championed in particular by founding chief executive Oren Etzioni and his successor Ali Farhadi, positions Ai2 alongside Hugging Face, EleutherAI, and to a lesser extent Mistral AI as a counterweight to the closed development practices of OpenAI, Anthropic, and Google DeepMind.[30] The institute's flagship language model series, OLMo, is the most thoroughly open foundation model family ever released at frontier scale.[12][13]
The Allen Institute for AI is a 501(c)(3) non-profit research institute whose stated mission is "AI research and engineering in service of the common good."[2] In practice that mission has three defining features. First, Ai2 conducts long-horizon research outside the product and time constraints of corporate labs and declines classified or military work.[2] Second, it publishes the substantial majority of its research in peer-reviewed venues with full reproducibility artifacts.[2] Third, and most distinctively, it ships its foundation models under a stricter definition of "open" than its peers: where Meta's Llama family, Mistral, and Google's Gemma release weights but withhold training data, code, or logs, Ai2 releases all of them.[13][14] On the OLMo 3 launch the institute summarized this stance plainly: "True openness in AI isn't just about access, it's about trust, accountability, and shared progress," arguing that "the models shaping our future should be fully inspectable, not black boxes."[41]
| Attribute | Detail |
|---|---|
| Type | 501(c)(3) non-profit research institute |
| Founded | January 2014 |
| Founder | Paul G. Allen |
| Headquarters | Seattle, Washington (Eastlake neighborhood) |
| Other offices | Tel Aviv, Israel |
| Annual operating budget | Approximately $50 to $80 million (mid-2020s); reported revenues and expenses each exceeding $110 million in some years |
| Employees | Roughly 200 to 250 researchers, engineers, and staff |
| Funding source | Endowment from the Paul G. Allen estate, channeled through the Vulcan Inc. (since renamed Vale Group) family-office structure |
| Mission | "AI research and engineering in service of the common good" |
| Founding CEO | Oren Etzioni (2013 to 2022) |
| Current CEO | Peter Clark (interim, 2026 to present), succeeding Ali Farhadi (2023 to 2026) |
| Website | allenai.org |
The institute is organized around three publicly stated focus areas: an Open Ecosystem program that produces fully open foundation models and shared infrastructure, an AI for Science program that builds tools to assist researchers in fields such as biomedicine and materials, and an AI for the Planet program that applies AI to conservation and environmental monitoring.[2] A fourth, smaller program covers embodied AI and robotics through the AI2-THOR simulator and downstream agent research.[40]
The Allen Institute for AI was founded in January 2014, when Oren Etzioni formally began as founding chief executive officer.[1][3] Paul Allen had recruited Etzioni, then a long-time professor at the University of Washington's Paul G. Allen School of Computer Science and Engineering, in September 2013 to direct research at the new organization.[3]
Paul Allen, who co-founded Microsoft with Bill Gates in 1975 and went on to become a major philanthropist after leaving the company, had a lifelong interest in artificial intelligence dating to his teenage years reading science fiction.[6][7] By the early 2010s he had already established the Allen Institute for Brain Science (founded 2003) and was exploring the idea of a similar institute focused on AI.[5] In September 2013 Allen recruited Oren Etzioni, then a long-time professor at the University of Washington's Paul G. Allen School of Computer Science and Engineering, to direct research at the new organization.[3] Etzioni took a leave of absence from his faculty position and formally began as founding chief executive officer in January 2014.[1][3]
The institute opened its doors that month with an initial endowment commitment from Allen reported in the press as exceeding $125 million, and a small founding team that included Peter Clark (formerly of Boeing's research division) and Carissa Schoenick.[1] The original mandate, broadly stated, was to undertake long-horizon AI research that fell outside the time and product constraints of corporate laboratories, with a particular early emphasis on machine reading, common-sense reasoning, and scientific question answering.[2]
The first project the institute publicly announced was Aristo, an effort to build an AI system capable of passing standardized science exams designed for human students.[10] A second early project, Semantic Scholar, applied natural language processing to the academic literature corpus to build a search engine for scientific papers.[24] Both projects were emblematic of the institute's interest in measurable, high-impact applications rather than narrow benchmark chasing.
Under Etzioni's leadership the institute grew from a few dozen people into one of the largest dedicated AI research labs outside of the major commercial laboratories.[1] By 2017 the institute reported a research staff of roughly 80 people and was producing peer-reviewed publications at a pace comparable to the most active university computer science departments.[32] By 2021 researchers at Ai2 had published nearly 700 papers, including a long line of contributions to top-tier venues such as NeurIPS, ICLR, ICML, ACL, NAACL, EMNLP, and CVPR.[1]
During this period the institute became closely associated with several breakthroughs in NLP. In 2018 a team led by Matthew E. Peters published Deep Contextualized Word Representations at the North American Chapter of the Association for Computational Linguistics conference, introducing ELMo, a deep bidirectional language model whose hidden states could be used as contextual word embeddings.[8][39] ELMo received the NAACL 2018 Best Paper Award and is widely regarded, alongside ULMFiT and the Transformer, as one of the works that inaugurated the transfer-learning era in NLP.[8][39]
The AllenNLP open-source library, released in 2017 and built on top of PyTorch, became one of the most widely used research toolkits for NLP experimentation.[1] AllenNLP shipped with implementations of ELMo, BiDAF, and a steady stream of other reference models, and it became the default starting point for many graduate students and researchers entering NLP through the late 2010s. The library was retired from active development at the end of 2022, with Ai2 redirecting its efforts toward foundation-model infrastructure that came to underpin OLMo and Tülu.[1]
In parallel, Project Aristo continued to mature. In 2019 the institute announced that Aristo had crossed an important threshold: on the non-diagram, multiple-choice portion of the eighth-grade New York Regents Science Exam the system scored 91.6 percent, comfortably above passing, and exceeded 83 percent on the twelfth-grade equivalent.[10][31] Just three years earlier the best AI system in a public competition had managed only 59.3 percent on similar material.[10] The Aristo paper, From 'F' to 'A' on the N.Y. Regents Science Exams, became a frequently cited milestone in the history of machine reading and reasoning.[10]
Etzioni stepped down as CEO on September 30, 2022 after almost nine years at the helm.[3] He continued to advise the institute as a board member and pivoted to entrepreneurship, founding Vercept and the deepfake-detection nonprofit TrueMedia.org, while remaining a venture partner at Madrona Venture Group.[3] Peter Clark served as interim chief executive during the search for a permanent successor.[29]
On July 31, 2023 Ali Farhadi became CEO.[4][29] Farhadi was already deeply tied to Ai2 and to the Seattle AI ecosystem; he is a professor at the University of Washington's Paul G. Allen School of Computer Science and Engineering and had previously co-founded XNOR.ai, an Ai2 Incubator startup acquired by Apple in 2020 for a reported $200 million.[4] He is married to Hannaneh Hajishirzi, the senior director of NLP research at Ai2, and the two had been collaborating on language-model research at the institute for years.[4]
Farhadi's appointment marked an unmistakable shift in the institute's external posture. Where the Etzioni years had been characterized by a portfolio of focused research projects, the Farhadi era reframed the institute around a single high-stakes thesis: that the scientific community urgently needed a fully open alternative to the increasingly closed frontier laboratories.[30] Under his leadership Ai2 launched OLMo in February 2024, the first frontier-class open language model released alongside its full pretraining data, training code, intermediate checkpoints, training logs, and evaluation harness.[12][13][14] The release was framed explicitly as an attempt to enable scientific reproducibility in language model research, a property that had largely vanished as commercial labs stopped publishing details of their systems.[13]
The institute followed OLMo with a rapid sequence of additional fully open releases. Tülu 2 (November 2023) and Tülu 3 (November 2024) demonstrated state-of-the-art post-training and instruction-tuning recipes.[19][20] Dolma (August 2023) provided a three-trillion-token open pretraining corpus.[21] OLMo 2 (November 2024) extended the family to 7B and 13B parameter models with substantially improved performance, and OLMo 2 32B (March 2025) became the first fully open model to outperform GPT-3.5 Turbo and GPT-4o mini on a battery of multi-skill benchmarks.[15][16][17] Molmo (September 2024) brought the same openness philosophy to multimodal vision-language models.[22] By the time OLMo 3 launched in November 2025, the institute was the unambiguous standard-bearer for transparent foundation-model research.[18][41]
In a February 2025 interview, after the open Chinese model DeepSeek R1 disrupted assumptions about closed-lab dominance, Farhadi framed openness as both a scientific and a national-competitiveness argument: "We're celebrating this moment as a proof point that open source will win," he said, adding that "if the U.S. wants to maintain its edge, we have only one way, and that is to promote open approaches, promote open-source solutions."[42]
Farhadi stepped down as CEO on March 12, 2026, after roughly two and a half years leading the institute, with co-founder Peter Clark again resuming the role of interim chief executive while the board searched for a permanent successor.[43][44] Board chair Bill Hilf told GeekWire that Farhadi wanted to pursue research at the extreme frontier of AI, where for-profit companies are spending billions training the most advanced models.[44] Shortly afterward Microsoft hired Farhadi, NLP research lead Hannaneh Hajishirzi, computer-vision researcher Ranjay Krishna, and former Ai2 chief operating officer Sophie Lebrecht to join Mustafa Suleyman's Superintelligence team, with the researchers retaining their University of Washington faculty appointments.[44] The departures were widely read as a stress test of whether a single non-profit lab can retain frontier talent against the compensation and compute of commercial superlabs.[43]
The institute's stated mission, repeated in nearly every public-facing communication, is to conduct "AI research and engineering in service of the common good."[2] Concretely this has meant several recurring commitments. First, Ai2 declines to take on classified or military research and does not pursue revenue-driven product development that would conflict with publishing. Second, the institute publishes the substantial majority of its research in peer-reviewed venues with full reproducibility artifacts.[2] Third, it releases its software and model weights under permissive open-source licenses.[13] Fourth, since the early 2020s it has prioritized work that other organizations are structurally unwilling or unable to undertake, with the OLMo program serving as the central example.[30]
In interviews following his appointment, Farhadi argued that the field of AI was at risk of becoming closed in a way that science had not been since the era before the internet.[30] He characterized OLMo and the broader Ai2 program as a deliberate attempt to keep the next generation of foundation models open enough to be studied, criticized, and improved by the global research community.[30] The argument resonated widely with academic researchers, government science agencies, and a growing community of independent AI safety and policy organizations who had become concerned that the most consequential AI systems were being developed inside black boxes.[30]
Ai2 is a 501(c)(3) non-profit organization.[36] Its endowment derives from Paul Allen's estate and is administered through the Vulcan Inc. family-office structure that Allen and his sister Jody Allen established in 1986.[5] (Vulcan was renamed Vale Group in 2024 as part of a broader reorganization of Allen family assets following Paul Allen's death in 2018.)[5][37] The institute's annual operating budget has been reported in the range of $50 million to $80 million in most years, although filings show revenues and expenses each exceeding $110 million during peak years that included large compute and infrastructure outlays for the OLMo program.[36]
Governance is overseen by a board of directors.[2] As of the mid-2020s the board has been chaired by Bill Hilf, the former Vulcan chief executive, with members including Jody Allen, University of Washington president emerita Ana Mari Cauce, Steve Hall, and computer science professor Ed Lazowska.[2] The board has historically allowed the chief executive substantial latitude in setting the research agenda, a pattern consistent with the broader Allen-family approach of recruiting strong scientific leaders and funding them at scale.
| Period | Chief executive | Background |
|---|---|---|
| 2014 to 2022 | Oren Etzioni (founding CEO) | Computer science professor at the University of Washington; previously founded MetaCrawler, Netbot (sold to Excite for $35 million), Farecast (acquired by Microsoft for $115 million), and Decide.com (acquired by eBay) |
| 2022 to 2023 | Peter Clark (interim) | Long-time AI2 senior research manager; Aristo project lead; previously at Boeing |
| 2023 to 2026 | Ali Farhadi | Professor at the Paul G. Allen School; co-founder of XNOR.ai (acquired by Apple in 2020); recipient of a Sloan Research Fellowship in 2017; departed March 12, 2026 to join Microsoft's Superintelligence team |
| 2026 to present | Peter Clark (interim) | Reprised the interim role after Farhadi stepped down |
Ai2 organizes its research into a small number of long-running programs and project areas, each of which has produced widely cited outputs. The table below summarizes the most prominent.
| Project | Year first released | Category | Description |
|---|---|---|---|
| Aristo | 2014 (project), 2019 (Regents milestone) | Reasoning and QA | Long-running project to build AI systems capable of passing standardized science exams; achieved over 90 percent on the eighth-grade New York Regents non-diagram multiple-choice questions in 2019 |
| Semantic Scholar | November 2015 | AI for Science | AI-powered academic literature search and discovery engine indexing more than 200 million papers; provides paper summaries, citation graph analysis, and Author Pages |
| AllenNLP | September 2017 | NLP infrastructure | Open-source PyTorch-based research library that became the standard NLP toolkit for many academic groups; retired at the end of 2022 |
| ELMo | February 2018 | Language modeling | Deep contextualized word representations from a bidirectional LSTM language model; NAACL 2018 Best Paper Award winner |
| ARC (AI2 Reasoning Challenge) | March 2018 | Benchmark | Question-answering benchmark of 7,787 grade-school science questions partitioned into Easy and Challenge sets; the Challenge set has become a standard rung on the modern LLM evaluation ladder |
| AI2-THOR | December 2017 | Embodied AI | Open-source 3D simulation framework built on Unity for training visual navigation and manipulation agents in photorealistic indoor scenes |
| Mosaic / ATOMIC / COMET | 2018 onward | Common-sense reasoning | Yejin Choi-led program that produced the ATOMIC knowledge graph (one million if-then social commonsense relations) and the COMET generative reasoning model |
| EarthRanger | 2017 onward (Ai2 stewardship) | AI for the Planet | Open-source platform integrating GPS collars, camera traps, and patrol reports for protected-area management; deployed across more than 900 sites in 95 countries |
| Dolma | August 2023 | Open data | Three-trillion-token English pretraining corpus released with full documentation, processing pipelines, and a permissive license; the foundation for OLMo |
| OLMo (1) | February 2024 | Foundation model | First fully open frontier language model, released with weights, data, training code, and evaluation; available at 1B and 7B parameter scales |
| Tülu 2 / Tülu 3 | November 2023 / November 2024 | Post-training | State-of-the-art instruction-tuning recipes built on Llama-family base models; Tülu 3 introduced Reinforcement Learning with Verifiable Rewards (RLVR) at scale |
| OLMo 2 | November 2024 | Foundation model | Second-generation OLMo at 7B and 13B parameters, trained on Dolmino data |
| Molmo | September 2024 | Multimodal model | Open vision-language model family trained on the PixMo image-text dataset; smaller variants approached GPT-4V on standard benchmarks |
| OLMo 2 32B | March 2025 | Foundation model | First fully open model to outperform GPT-3.5 Turbo and GPT-4o mini on a suite of multi-skill academic benchmarks; trained on roughly 6 trillion tokens |
| OLMo 3 | November 2025 | Foundation model | Family of frontier-class open models including the first fully open 32B reasoning model that produces explicit chain-of-thought style content |
| Molmo 2 | December 2025 | Multimodal model | Extended Molmo to video understanding, pointing, and tracking |
Aristo, named for Aristotle, was the institute's flagship reasoning project from 2014 through the early 2020s.[10] The system tackled science exam questions written for human students, beginning with grade-school multiple choice and progressing through high-school exam material.[10] Early versions relied on hand-engineered solvers that combined information retrieval, statistical association, structured knowledge bases, and rule-based reasoning. Later versions incorporated transformer-based language models.[10] The September 2019 announcement that Aristo had crossed 90 percent on the eighth-grade New York Regents non-diagram multiple-choice questions was widely covered in the technology press as a sign that machine reading was finally making concrete progress on tasks that had stymied AI for decades.[31] The arXiv paper From 'F' to 'A' on the N.Y. Regents Science Exams documents the system in detail.[10] Project Aristo continues today as part of the institute's broader reasoning agenda, with current emphasis on systematic explanation generation and verifiable reasoning chains.
Semantic Scholar, launched in November 2015, is a free academic search engine that uses NLP and machine learning to surface, summarize, and recommend research papers.[24][25] It indexes more than 200 million papers across nearly every scientific discipline, with deep coverage of computer science, biomedicine, neuroscience, and physics.[24] Distinctive features include AI-generated paper summaries (TL;DRs), automatic citation parsing, the Citation Graph API, Author Pages, the Specter document embedding model, and personalized recommendations.[24][25] Semantic Scholar's underlying corpus, the Semantic Scholar Open Research Corpus (S2ORC), has become a heavily used dataset in scientific NLP research and has been incorporated into the pretraining mixes of many open language models.[25] The product is now part of the broader AI for Science portfolio at Ai2, which also includes the Asta scientific assistant project and the SciREX information extraction effort.
The AllenNLP library, first released in September 2017, was for several years the dominant open-source research framework for natural language processing.[1] It was built on PyTorch and shipped with reference implementations of leading models, an experiment-tracking framework, and a permissive Apache 2.0 license. The library played an outsized role in democratizing modern NLP research because it lowered the barrier for graduate students and independent researchers to reproduce state-of-the-art results.
The single most influential model released through the AllenNLP infrastructure was ELMo, short for Embeddings from Language Models.[39] The original paper, Deep Contextualized Word Representations by Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer, appeared at NAACL 2018 and won the conference's Best Paper Award.[8][39] ELMo trained a deep bidirectional LSTM as a language model on the One Billion Word Benchmark and then used a learned linear combination of the LSTM's internal hidden states as task-agnostic word representations.[8] Models that previously relied on static GloVe or Word2Vec embeddings saw immediate accuracy gains across most NLP benchmarks when ELMo was substituted in.[8] The technique helped popularize the idea that pretrained language models, used as feature extractors or fine-tuned for downstream tasks, would become the default substrate for NLP, a trajectory that ran straight through GPT, BERT, and the entire modern foundation-model era.[39]
AllenNLP was officially retired at the end of 2022, with the team redirecting effort toward the open-instruct, OLMo, and dolma codebases that now form the backbone of the institute's open foundation-model program.[1]
In March 2018 a team including Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord released Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge.[9] The accompanying dataset became one of the standard hard reasoning benchmarks in the field.[9] ARC contains 7,787 grade-school science questions split into an Easy Set and a Challenge Set; the Challenge Set was specifically constructed to include only questions that defeated both information-retrieval and word-co-occurrence baselines, leaving items that genuinely require multi-step reasoning.[9] Modern large language model evaluations almost invariably report ARC-Challenge accuracy alongside MMLU, HellaSwag, TruthfulQA, GSM8K, and similar benchmarks. The dataset is hosted on the Hugging Face Hub and remains in active use almost a decade after release.[9]
The Perceptual Reasoning and Interaction Research team, known internally as PRIOR, focuses on computer vision, embodied AI, and interactive agents. Its flagship release was AI2-THOR, an open-source 3D simulation framework first published in December 2017.[11][40] AI2-THOR uses the Unity 3D engine to render physically plausible indoor scenes, including kitchens, bedrooms, living rooms, and bathrooms, in which agents can navigate, manipulate objects, open containers, and interact with simulated appliances.[11][40] The framework supports tasks ranging from visual question answering to instruction following to object goal navigation.[11] Subsequent releases extended the platform with ProcTHOR (procedural scene generation), ManipulaTHOR (richer manipulation), RoboTHOR (real-to-sim transfer for physical robots), and ObjectNav benchmarks.[40] AI2-THOR has been cited in thousands of papers and is widely used by academic groups working on embodied agents.[40]
The Mosaic team, founded around 2018 and led by Yejin Choi (then a UW professor and Senior Director of Mosaic at Ai2), pursued the long-standing problem of giving machines common sense.[38] The team produced the ATOMIC commonsense knowledge graph, which encodes more than one million if-then social inferences (for example, if X pays Y a compliment, Y will likely return the compliment), and the COMET generative model, which extends symbolic knowledge graphs by generating plausible new commonsense inferences for events not in the underlying graph.[19] Other Mosaic outputs included the Social Bias Frames dataset, the Defeasible NLI benchmark, the Delphi moral judgment system, and a long line of papers on social reasoning and value alignment. Choi was awarded a MacArthur Fellowship in 2022 and left Ai2 for Stanford University in 2024.[38]
The AI for the Planet program at Ai2 builds open-source software and machine-learning systems for ecosystem monitoring and conservation.[27] Its flagship product is EarthRanger, an open-source platform that integrates real-time data from GPS animal collars, camera traps, ranger patrol reports, satellite feeds, and environmental sensors into a unified interactive map and alerting system.[26] EarthRanger originated as a Vulcan project sponsored by Allen's earlier conservation work and was transferred to Ai2's stewardship as the institute expanded its planetary program.[26] As of the mid-2020s the platform was deployed across more than 900 protected areas in 95 countries, tracking more than 23,000 animals and supporting park rangers, conservation NGOs, and government agencies.[26] The institute provides hosting, deployment, and ongoing development at no cost to qualifying organizations.[26] Sister projects in the AI for the Planet portfolio include Skylight (an illegal-fishing detection platform), OlmoEarth (an environmental modeling effort), and various climate and biodiversity research programs.[27]
The institute's most visible work since 2023 has been its open foundation-model program.[12] The program operates under the assertion that a frontier AI model is not truly open unless its weights, training data, training code, intermediate checkpoints, training logs, and evaluation harness are all available under permissive licenses.[13] By that standard most so-called "open" releases from major labs (including the Llama family from Meta, Mistral models, and Gemma from Google) are partially open at best, since they release weights but not data, code, or logs.[14] OLMo and its sibling releases were designed from the start to meet the stricter definition.[12][13]
Yes. Ai2 is generally regarded as the most committed of any frontier-scale lab to fully open release, and it applies a deliberately strict test: a model counts as open only if its weights, pretraining data, training code, intermediate checkpoints, training logs, and evaluation harness are all published under permissive licenses.[13][14] On the OLMo 2 32B release the institute described its goal as packaging everything into "a single, turn-key recipe with all ingredients openly available: data, training recipe, methodology, and software."[45] This standard is stricter than the open-weight releases of Meta's Llama, Mistral, or Google's Gemma, which publish weights but withhold significant portions of the data and training pipeline.[14]
Dolma, first released in August 2023, is an open three-trillion-token English-language pretraining corpus drawn from web pages, scientific publications, code repositories, books, encyclopedic text, and social-media-adjacent text.[21] The dataset is distributed with full documentation of its sources, processing scripts, decontamination routines, and license terms.[21] Dolma 2 followed in late 2024 with improved quality filtering, and Dolma 3, a roughly 9.3-trillion-token corpus whose 5.9-trillion-token pretraining mix powered OLMo 3, accompanied that release in November 2025.[18][41] Dolma was deliberately designed so that researchers could examine, modify, or rebuild any portion of it; the toolkit for constructing Dolma is itself open source.[21]
The OLMo series is the flagship of Ai2's open foundation-model program.[12] Each release expands the family across parameter counts, training-token budgets, and capability axes while preserving the strict openness criterion.[13]
| Release | Date | Parameter scales | Training tokens | Notes |
|---|---|---|---|---|
| OLMo 1 | February 1, 2024 | 1B and 7B | Approximately 2.5 trillion (Dolma) | First fully open frontier language model; released with full data, code, weights, checkpoints, and training logs |
| OLMo 1.7 | July 2024 | 7B | Approximately 2.5 trillion | Refresh with improved data mix and training procedure |
| OLMo 2 | November 2024 | 7B, 13B | Approximately 5 trillion | Trained on Dolmino-mix data; substantially better evaluation results than OLMo 1 |
| OLMo 2 32B | March 13, 2025 | 32B | Approximately 6 trillion | First fully open model to outperform GPT-3.5 Turbo and GPT-4o mini on multi-skill academic benchmarks; post-trained with Tülu 3.1; reported to require roughly one third the training cost of comparable Qwen 2.5 32B |
| OLMo 3 | November 20, 2025 | 7B and 32B (Base, Think, plus 7B Instruct and RL Zero) | Approximately 5.9 trillion (Dolma 3 Mix) | Introduced the "model flow" concept and the first fully open 32B "thinking" model with explicit reasoning-chain content; OLMo 3-Think 32B is described as the strongest fully open thinking model to date; built with the Dolci post-training stack |
The OLMo family is hosted on the Hugging Face Hub, with weights, intermediate checkpoints (at thousands of training steps), training scripts, evaluation suites, and write-ups available.[12][13] The codebase is maintained at github.com/allenai/OLMo.[12] Training was performed in collaboration with Databricks, AMD, the Paul G. Allen School at the University of Washington, and the Kempner Institute at Harvard University.[14]
OLMo 3, announced on November 20, 2025 and documented in a paper with more than 60 Ai2 authors (arXiv:2512.13961), is built around the concept of the "model flow": Ai2 defines this as "the full lifecycle of a language model: every stage, checkpoint, dataset, and dependency required to create and modify it," released so that anyone can intervene at any point rather than only fine-tune final weights.[41][46] The family spans 7B and 32B scales and includes four 7B variants (Base, Instruct, Think, and RL Zero) plus 32B Base and 32B Think models.[41][46] OLMo 3-Think 32B is, according to Ai2, the strongest fully open reasoning model released to date, competitive on math and reasoning benchmarks with Qwen 3 VL 32B Thinking and the DeepSeek R1 distillation at the same scale.[41] The release reinforced the institute's central claim that "the models shaping our future should be fully inspectable, not black boxes."[41]
The Tülu series, named after a hybrid camelid, packages Ai2's evolving recipes for post-training open language models.[19] Tülu 1 (June 2023) and Tülu 2 (November 2023) demonstrated competitive instruction-tuned models built on Llama 1 and Llama 2 base weights respectively.[19] Tülu 3 (November 2024) was a more ambitious release: it provided a complete open recipe for state-of-the-art post-training, including supervised fine-tuning on a curated mix of public, synthetic, and human data, Direct Preference Optimization (DPO) on a tens-of-thousands preference set, and a novel Reinforcement Learning with Verifiable Rewards (RLVR) stage.[19][20] The Tülu 3 codebase, dataset mixes, and evaluation framework were all released openly.[19][20] The same recipe, refined into Tülu 3.1 (which adds Group Relative Policy Optimization to the RLVR stage), underpins the OLMo-Instruct variants of OLMo 2 and OLMo 3.[16][45]
In September 2024 Ai2 released Molmo, an open family of vision-language models.[22] Where most multimodal model releases at the time disclosed weights but not the underlying training data, Molmo was released alongside PixMo, the curated training dataset of just under one million image-text pairs that powered the family.[22] The smallest variant, MolmoE-1B (built on the OLMoE-1B-7B mixture-of-experts base), nearly matched GPT-4V on standard multimodal benchmarks.[22] Larger variants, including Molmo-7B-D and Molmo-72B, scored competitively against GPT-4V and approached GPT-4o on academic benchmarks and human preference evaluations.[22] A distinctive Molmo capability was point grounding: the model learned to point at specific image regions in response to queries, enabling more precise interactions with physical and virtual environments.[22] Molmo 2, released in December 2025, extended the program to video understanding, multi-frame reasoning, and tracking.[23]
Ai2 has employed or hosted many of the most prolific researchers in modern NLP, computer vision, and reasoning. The table below lists a selection of researchers most closely associated with the institute.
| Researcher | Role at Ai2 | Notable contributions |
|---|---|---|
| Oren Etzioni | Founding CEO (2014 to 2022); board member | Founded the institute; led the Aristo program in its early years; previously coined the term machine reading |
| Peter Clark | Senior Research Manager; Aristo lead; interim CEO (2022 to 2023, 2026 to present) | Aristo program lead; first author of the F to A on the N.Y. Regents Science Exams paper; co-author of ARC |
| Ali Farhadi | CEO (2023 to 2026); UW professor | Co-founder of XNOR.ai; led the institute's pivot to fully open foundation models; left for Microsoft's Superintelligence team in 2026 |
| Hannaneh Hajishirzi | Senior Director of NLP Research; UW professor | Long-time leader of NLP research at Ai2; co-led the OLMo and Tülu programs; left for Microsoft's Superintelligence team in 2026 |
| Yejin Choi | Senior Director of Mosaic (until 2024); UW professor (until 2024); now at Stanford | Mosaic program lead; ATOMIC, COMET, Delphi, Social Bias Frames; MacArthur Fellow 2022 |
| Doug Downey | Senior Research Manager; Northwestern professor | Semantic Scholar, search and recommendation systems, scientific NLP |
| Matthew E. Peters | Research Scientist (formerly) | First author of the original ELMo paper |
| Matt Gardner | Research Scientist (formerly) | AllenNLP architect; ELMo co-author; later co-founded a startup |
| Mark Neumann | Research Scientist (formerly) | ELMo co-author; key AllenNLP contributor |
| Luke Zettlemoyer | Research Scientist; UW professor | ELMo co-author; long history of contributions to semantic parsing and language modeling |
| Iz Beltagy | Research Scientist | SciBERT, Longformer, and other long-context model work |
| Kyle Lo | Research Scientist | Semantic Scholar; S2ORC corpus; Dolma |
| Pradeep Dasigi | Research Scientist | Reading comprehension and reasoning evaluation |
| Tushar Khot | Research Scientist | ARC; chain-of-thought reasoning research; decomposition prompting |
| Roy Schwartz | Research Scientist (formerly) | Energy-efficient NLP and the Green AI agenda |
| Roozbeh Mottaghi | Research Manager (formerly) | AI2-THOR, embodied AI |
| Aniruddha Kembhavi | Research Manager | Multimodal reasoning, AI2-THOR, Molmo |
| Eric Kolve | Engineering Manager | AI2-THOR engineering lead |
Many Ai2 researchers hold joint appointments with the University of Washington's Paul G. Allen School of Computer Science and Engineering, reflecting the institute's deliberate co-location next to one of the world's leading academic AI departments.[1] Others have come from or moved to peer institutions including Carnegie Mellon University, MIT, Stanford University, Cornell University, the University of Edinburgh, and Northwestern University.
The AI2 Incubator was established in 2017 as an in-house program to spin out AI startups built around the institute's research.[34] It operated under the institute's umbrella for several years and in 2022 was reorganized as a separate legal entity, although it retained close ties to Ai2 and continued to draw on its research community.[33] The incubator has spun out more than forty companies with combined value in the billions of dollars; notable exits include XNOR.ai (acquired by Apple in 2020 for a reported $200 million), Lexion (acquired by DocuSign in 2024), and KITT.AI (acquired by Baidu in 2017).[33][34] Other portfolio companies have included WellSaid Labs, Yellowdig, Velocity AI, and Soundsensing. The incubator's first dedicated investment fund of $10 million closed in 2020, followed by a $30 million Fund II in 2023 and an $80 million Fund III in 2025 backed by Madrona Venture Group, SCB 10X, and other strategic limited partners.[33][34]
Ai2 occupies a distinctive position in the global AI research landscape. It is one of a small handful of organizations capable of training frontier-scale foundation models that is neither a commercial laboratory nor a national security program, and it is the most committed of those organizations to the principle of fully open release.[30] Comparable organizations include:
Where OpenAI and Anthropic operate as venture-backed, profit-oriented laboratories that release neither weights nor data for their frontier systems, Ai2 has positioned itself as the institutional standard-bearer for the proposition that scientific progress in AI requires unrestricted reproducibility.[30] The contrast has become more pronounced as the gap between proprietary and open frontier models has narrowed and as government science funders have sought open foundation-model programs to support broader research.
Ai2 is also a participant in the National AI Research Resource (NAIRR) Pilot, the United States federal government's program to provide academic researchers with shared access to computational resources and data.[35] The institute committed in November 2023 to contribute infrastructure, datasets, and models to the program, in line with its general posture of subsidizing the open ecosystem.[35]
The central difference is openness. OpenAI and Anthropic are venture-backed, for-profit laboratories that release neither the weights nor the training data of their frontier systems, treating those assets as proprietary.[30] Ai2 is a non-profit that releases the complete pipeline, weights, data, code, checkpoints, and logs, under permissive licenses, on the argument that reproducibility is a precondition for scientific progress in AI.[13][30] A second difference is scale and intent: Ai2's stated aim is not to win every benchmark but, in Farhadi's framing, to ensure that fully reproducible models exist within striking distance of the frontier so that independent research can continue.[30][42]
The following papers from Ai2 researchers have had outsized impact on the field. The list is far from exhaustive.
| Year | Paper | Lead authors | Significance |
|---|---|---|---|
| 2018 | Deep Contextualized Word Representations (ELMo) | Matthew E. Peters et al. | NAACL 2018 Best Paper Award; introduced contextual word embeddings |
| 2018 | Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge | Peter Clark et al. | Introduced the ARC benchmark |
| 2019 | From 'F' to 'A' on the N.Y. Regents Science Exams | Peter Clark et al. | Documented the Aristo project's eighth-grade exam milestone |
| 2019 | AllenNLP Interpret | Eric Wallace et al. | Open-source NLP model interpretation framework |
| 2019 | ATOMIC: An Atlas of Machine Commonsense for If-Then Reasoning | Maarten Sap, Yejin Choi et al. | Million-edge commonsense knowledge graph |
| 2019 | SciBERT: A Pretrained Language Model for Scientific Text | Iz Beltagy et al. | Specialized BERT variant for science papers |
| 2020 | Longformer: The Long-Document Transformer | Iz Beltagy et al. | Sparse attention for long context |
| 2023 | Dolma: an Open Corpus of Three Trillion Tokens | Luca Soldaini et al. | Open pretraining corpus |
| 2024 | OLMo: Accelerating the Science of Language Models | Dirk Groeneveld et al. | First fully open frontier language model |
| 2024 | Tülu 3: Pushing Frontiers in Open Language Model Post-Training | Nathan Lambert et al. | Open post-training recipes including RLVR |
| 2024 | Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Multimodal Models | Matt Deitke et al. | Open multimodal vision-language models |
| 2025 | OLMo 2: A Family of Open Language Models | OLMo team | Second-generation open foundation models |
| 2025 | Olmo 3 (arXiv:2512.13961) | Olmo 3 team (60+ authors) | Frontier open model family introducing the model flow and OLMo 3-Think 32B |
By the mid-2020s Ai2 had become one of the most influential voices in the global debate over how foundation models should be developed and governed.[30] Where commercial laboratories tended to argue that the safest path was tight control over weights and information, Ai2 leadership argued in policy forums that openness was essential both for scientific progress and for the diffusion of safety research itself.[30] In numerous Senate, House, and executive-branch consultations, the institute pressed for policies that preserved space for open foundation models, including in the National Telecommunications and Information Administration's 2024 review of dual-use foundation models with widely available weights, the Biden administration's 2023 executive order on AI, and similar policy processes in the European Union and the United Kingdom.
The institute has also been a significant employer and trainer of AI talent in the Pacific Northwest.[32] Many Ai2 alumni have gone on to senior research roles at OpenAI, Anthropic, Google DeepMind, Microsoft Research, Meta AI, and Apple, as well as to faculty positions across the United States, Europe, and Asia. The 2026 departure of CEO Ali Farhadi, NLP research lead Hannaneh Hajishirzi, and other senior staff to Microsoft's Superintelligence team underscored how acutely commercial superlabs compete for the institute's talent.[44] The institute's tight integration with the University of Washington has made Seattle one of the world's leading hubs for AI research, alongside the Bay Area, Boston, London, and the Greater Toronto Area.[1]
Like all major AI laboratories, Ai2 has been the subject of criticism on several fronts. Some commentators have argued that fully open release of capable foundation models is at odds with prevailing AI safety norms and could enable misuse; the institute has generally responded that the marginal risk from any individual open model is small relative to the safety benefits of broad transparency.[30] Others have noted that even Ai2's openness commitments depend on a large endowment underwritten by a single donor family, raising long-term sustainability questions if the Allen estate's priorities were ever to change.[36] The 2024 reorganization of Vulcan into Vale Group and the 2025 launch of the $3.1 billion Fund for Science and Technology by the Allen estate have generally been read as reaffirmations of long-term commitment to the institute, but the question of multi-decade sustainability remains live.[37]
A second line of criticism has focused on whether OLMo and similar releases meaningfully close the capability gap with the proprietary frontier. While OLMo 2 32B's claim to surpass GPT-3.5 Turbo and GPT-4o mini was widely reported, the very largest commercial models (GPT-4o, Claude Opus, Gemini Ultra) remained substantially ahead on most benchmarks throughout 2024 and into 2025.[16][17] Ai2 leadership has acknowledged the gap and argued that the goal is not to win every benchmark but to ensure that fully reproducible models exist within striking distance of the frontier so that scientific work can continue.[30]
A third challenge surfaced in 2026, when the near-simultaneous departure of CEO Ali Farhadi and several senior researchers to Microsoft raised questions about whether a single non-profit lab can retain frontier-level talent against the compensation and compute budgets of commercial superlabs.[43][44]