What Allen Institute for Artificial Intelligence does

AI2 (Allen Institute for Artificial Intelligence) is a Seattle-based 501(c)(3) non-profit AI research institute founded in 2014 by the late Microsoft co-founder Paul Allen. The institute operates as a fully open foundation-model platform and applied research organization, serving AI researchers, ML developers, scientific institutions, conservation organizations, and government agencies. Its work spans language, multimodal, embodied/robotics, and Earth-observation AI, with applied platforms for wildlife and ocean conservation alongside scientific discovery tooling.

The institute's product portfolio includes the Olmo family of fully open large language models (Olmo 3, Olmo 3.1, Tülu 3, Olmo Hybrid), the Molmo multimodal and vision-language models (Molmo 2, MolmoPoint, Unified-IO-2), embodied and robotics systems (MolmoAct 2, MolmoBot, MolmoSpaces), scientific AI agents and assistants (Asta, DR Tulu, OpenScholar, ScholarQA, AutoDiscovery), and Earth-observation and conservation platforms (OlmoEarth Platform, EarthRanger, Skylight, ACE climate emulators). The technology stack is anchored on end-to-end open model flows that expose training data, training code, intermediate checkpoints, and evaluation recipes alongside released weights, in contrast to closed commercial frontier labs. Underlying infrastructure includes the OMAI computing cluster built on NVIDIA Blackwell Ultra chips and managed by Cirrascale, supported by Ai2-developed training frameworks (OlmoCore, Open Instruct) and evaluation suites (OLMES, AstaBench). Large-scale open datasets (Dolma, S2ORC with 10M+ academic documents, S2AG with 200M+ paper metadata, WildChat, PixMo, Molmo 2 datasets) feed pretraining and benchmarking.

AI2's business model is grant- and philanthropy-driven rather than revenue-driven. Primary funding flows from the Paul G. Allen Trust, the Fund for Science and Technology, and major public-private grants including a $152M OMAI award from NSF ($75M) and NVIDIA ($77M). Most artifacts (models, data, code, weights) are released fully open-source under permissive licenses, generating adoption through community-led distribution (Hugging Face, GitHub, the Ai2 Playground). Commercial revenue streams are early-stage: the OlmoEarth Platform offers enterprise API access via request-account onboarding, with usage-based pricing in beta through June 2026. An affiliated AI2 Incubator manages an $80M fund for AI startups built on Ai2 research.

Name

Allen Institute for Artificial Intelligence

Legal name

The Allen Institute for Artificial Intelligence

Website

https://allenai.org

Company type

Private

Founded year

2014

Operating status

Operating

Headcount range

101–250 employees

Short description

Seattle-based 501(c)(3) non-profit AI research institute founded in 2014 by Paul Allen, developing open foundation models (Olmo, Molmo), scientific AI agents (Asta, DR Tulu), and Earth-observation platforms (OlmoEarth, EarthRanger) for researchers, conservation organizations, and government agencies.

Ownership category

akta.pro rank

Product category

Open AI Research & Foundation Models

NAICS

Software Publishers (513210), Computer Systems Design and Related Services (5415)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

End-to-End Enterprise AI Platforms (MLOps & Model Lifecycle Management) (HDAEANAA)

akta.pro secondary industries

AI Application Enablement Platforms (Copilot/Agent Frameworks, SDKs) (HDAEANAJ), Agents & Autonomous Workflows (Tool Use, Planning, Multi-Agent) (HDAAACAF), AI Observability, Monitoring & Evaluation Platforms (Drift, Quality, Safety) (HDAEANAF), Responsible AI, AI Governance & Compliance Services (BPAEAHAJ)

Keywords

• Open-source AI models
• Scientific AI research
• AI for the planet
• Embodied AI robotics
• Nonprofit AI institute

Headquarters

HQ city

Seattle

HQ country

United States

HQ region

North America

Offices3 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations, Marketing or Sales, Others

Revenue model

1. Non-profit grants and endowment funding: As a Seattle-based non-profit AI research institute founded by Paul Allen, Ai2's primary funding comes from philanthropic endowments and grants, including the Paul G. Allen Trust, Allen Family Philanthropies, and the Fund for Science and Technology, which provides proposal-based research funding.
2. Government and industry grants: Major grant funding from the National Science Foundation ($75 million) and NVIDIA ($77 million) for the $152 million Open Multimodal AI Infrastructure to Accelerate Science (OMAI) project, including a dedicated computing cluster on Nvidia Blackwell Ultra chips managed by Cirrascale.
3. OlmoEarth Platform commercial access: Enterprise access to the OlmoEarth Platform via request account, with API documentation and onboarding for organizations using Earth observation models for monitoring, mapping, and decision-making; includes free trial through June 2026.
4. Cloud partnership infrastructure (Google Cloud): Strategic partnership with Google Cloud supports development of Ai2's open AI models and AI platforms for scientific research and environmental modeling, providing infrastructure for model training and deployment.
5. AI2 Incubator $80M fund: The AI2 Incubator operates an $80M fund investing in and supporting AI startups, generating returns that can be reinvested in the institute's research mission.

Pricing tiers

Go-to-market motion4 records

Distribution channels7 records

Marketing channels13 records

Core offering

Ai2 develops and releases fully open foundation AI models (language, multimodal, robotics, Earth observation) along with end-to-end model flows exposing data, training code, weights, and recipes. On top of these models, Ai2 builds applications for science (Asta agent ecosystem, Semantic Scholar, ScholarQA, OpenScholar, DR Tulu, AutoDiscovery), for the planet (OlmoEarth, EarthRanger, Skylight, ACE climate emulators), and for embodied AI (MolmoAct, AI2-THOR, MolmoSpaces). Most offerings are released under permissive open-source licenses, with select commercial enterprise access (e.g., OlmoEarth Platform) via request-account onboarding.

Product overview

Ai2 (Allen Institute for Artificial Intelligence) is a Seattle-based non-profit AI research institute founded in 2014 by the late Paul Allen. Its portfolio is organized as an open foundation-models platform plus a portfolio of applications and tools. At the core, the Olmo family (Olmo 3, Olmo 2, Olmo Hybrid, OLMoE) and Tülu 3 provide fully open language models with end-to-end model flow transparency (Dolma data, OlmoCore training, Open Instruct post-training, OLMES evaluation), while the Molmo family (Molmo 2, MolmoPoint, MolmoWeb, MolmoAct, MolmoAct 2, MolmoBot, MolmoSpaces, Unified-IO/-IO-2, VisProg, GraspMolmo) extends open multimodal and embodied AI capabilities across image, video, audio, and robotics. On top of these foundation models, Ai2 builds applications for science (Asta agentic ecosystem with AstaBench, ScholarQA, Semantic Scholar, DR Tulu, OpenScholar, AutoDiscovery, olmOCR), for the planet (OlmoEarth Platform and foundation models, EarthRanger, Skylight, ACE/SamudrACE/FloeNet/HiRO-ACE climate emulators, Satlas, Wildlands), and for embodied AI (MolmoAct/MolmoAct 2 manipulation, MolmoBot, AI2-THOR, ProcTHOR, Objaverse/Objaverse-XL, Holodeck, GraspMolmo, RING, PoliFormer, SPOC, FLaRe). Cross-cutting infrastructure includes the Ai2 Safety Toolkit (WildTeaming, WildJailbreak, WildGuard), DataDecide, OLMES, the Ai2 Playground, and partner integrations for deployment.

Differentiator

Problem solved

Functional benefit

Brands

• Olmo: Fully open language model family and complete model flow from data to deployment, including Olmo 2 and Olmo 3 variants.
• Tülu (Tülu 3)
• Molmo
• OlmoEarth
• EarthRanger
• Skylight
• Asta
• Semantic Scholar
• Satlas
• ACE (Ai2 Climate Emulator)
• AI2-THOR

Products and services

• Olmo 3 Family of fully open language models offering full transparency of the model flow from data to deployment, with 32B and 7B Base, Think, and Instruct variants covering pretraining, mid-training, long-context, instruction SFT/DPO/RL, and RL Zero stages.
• Olmo 2 Family of 7B, 13B, and 32B language models trained up to 6T tokens, outperforming contemporaneous fully open models and competing with open-weight models like Llama 3.1 8B.
• Olmo Hybrid 7B base model combining transformer attention with gated delta net (GDN) recurrent layers, demonstrating approximately 2X pretraining efficiency gains and superior long-context performance.
• OLMoE First mixture-of-experts model in the OLMo family, on the Pareto frontier of performance and size, trained 2x faster than equivalent dense models, released with open data, code, evaluations, logs, and intermediate training checkpoints.
• Tülu 3 Leading open-source instruction-following model family with open data, code, and recipes for SFT, DPO, and RLVR; targets knowledge, reasoning, math, coding, instruction following, chat, and safety; built on Llama 3 Base.
• Molmo 2 Family of open state-of-the-art multimodal AI models in 4B, 8B, and O-7B variants expanding original Molmo strengths in grounded vision to video and multi-image understanding, including object tracking, event localization, captioning, pointing, and counting.
• Molmo (original) Family of open-source multimodal language models with 72B and 7B variants; reportedly outperforms GPT-4o on image, chart, and document understanding benchmarks; trained on a 600K-image curated dataset with detailed human descriptions.
• MolmoPoint Pointing architecture for vision-language models replacing text-coordinate generation with direct visual token selection using three special tokens (PATCH, SUBPATCH, LOCATION); available as MolmoPoint-8B, MolmoPoint-GUI-8B, and MolmoPoint-Vid-4B variants.
• MolmoWeb Open-source visual AI agent that controls web browsers and automates web tasks including navigating, clicking, typing, and interacting within browsers; available in 4B and 8B parameter sizes and trained with synthetic and human web browsing data.
• MolmoAct 2 Fully open robotics foundation model for real-world manipulation built on Molmo 2-ER with action expert and adaptive depth reasoning; substantially outperforms proprietary robotics models and handles real-world tasks faster than its predecessor; paired with the 720+ hour MolmoAct 2-Bimanual YAM dataset.
• MolmoAct First Action Reasoning Model that reasons in 3D space for robotics, built on the Molmo vision-language family; achieves 72.1% OOD task success on SimplerEnv and 86.6% on LIBERO; released fully open source with training code, checkpoints, and a post-training dataset of ~10,000 robot episodes.
• MolmoSpaces Fully open ecosystem for embodied research combining 230,000+ indoor scenes, 130,000+ curated object assets, and 42 million physics-grounded grasp annotations in a modular platform compatible with MuJoCo, ManiSkill, NVIDIA Isaac Lab, and NVIDIA Isaac Sim.
• AI2-THOR Open-source simulation platform for embodied AI and robotics research providing near photo-realistic 3D indoor environments in which virtual agents can navigate, interact with objects, and learn from their actions.
• Asta Agentic ecosystem for scientific discovery uniting Asta agents (research assistants for complex scholarly tasks), AstaBench (rigorous evaluation framework for AI agents), and Asta resources (tools, baseline agents, templates, and APIs).
• AstaBench Benchmark suite and leaderboards for evaluating AI agents on scientific tasks, comprising thousands of scientific research problems across literature understanding, code execution, data analysis, and end-to-end discovery categories.
• DR Tulu (Deep Research Tulu) First fully open, end-to-end training recipe for building deep research AI agents, including the DR Tulu-8B model checkpoint, training code, and supporting libraries; matches or outperforms Gemini 3 Pro and OpenAI Deep Research on ScholarQA-CSv2 at a fraction of the cost.
• OpenScholar Fully open retrieval-augmented language model that synthesizes citation-backed responses from 45 million open-access scientific papers; OpenScholar-8B outperforms GPT-4o by 6.1% in correctness while being a smaller open model.
• Semantic Scholar AI-driven scientific literature search and discovery platform helping researchers find relevant papers, explore new knowledge, stay up-to-date, and understand papers, backed by the S2ORC and S2AG datasets.
• Ai2 ScholarQA Scientific QA system answering questions that require multiple scientific papers, delivering in-depth contextual answers with table comparisons, expandable subtopics, and citations with paper excerpts for verification.
• AutoDiscovery AI-powered system available in AstaLabs that autonomously generates its own hypotheses, runs experiments, and uses Bayesian surprise to chase the findings most likely to represent genuine discoveries, compressing weeks of exploratory analysis into hours.
• olmOCR Open-source tool designed for high-throughput conversion of PDFs and other documents into plain text while preserving natural reading order, supporting tables, equations, handwriting, and more; backed by olmOCR-bench with 1,403 PDF files and 7,010 unit test cases.
• OlmoEarth Platform End-to-end solution for scalable planetary intelligence integrating foundation models with planetary-scale datasets and adaptable infrastructure, turning Earth data into decision-ready insights for conservation, wildfire risk, mangrove mapping, and ocean monitoring.
• EarthRanger Flexible, powerful technology platform that empowers protected area managers, ecologists, and wildlife biologists across the globe to make more informed operational decisions for wildlife conservation through real-time monitoring, sensors, and AI-connected devices.
• Skylight AI-powered ocean monitoring platform leveraging state-of-the-art AI to detect illegal, unreported, and unregulated (IUU) fishing, processing 100M+ daily GPS data points and satellite imagery from NASA, ESA, and Maxar.
• Ai2 Climate Emulator (ACE) Fast, accurate, state-of-the-art open-source AI emulator of weather and climate trained on observational reanalysis such as ERA5 or a physics-based climate model; trains on global storm-resolving models to emulate similar climates thousands of times faster.
• Unified-IO 2 First autoregressive multimodal model capable of understanding and generating images, text, audio, and action; achieves state-of-the-art on the GRIT benchmark and strong results across more than 30 benchmarks.
• Ai2 Playground Hosted environment at playground.allenai.org where users can try Olmo, Tülu 3, and Molmo models via a chat interface.
• Satlas Tool using satellite imagery to show how Earth's surface is changing, with AI models enhancing low-resolution satellite imagery to produce high-resolution images on a global scale.
• CoSyn Tool developed with Penn Engineering that uses open-source AI models to generate synthetic scientific figures, charts, and tables for training other AI systems to interpret complex visual information; CoSyn-trained models match or outperform GPT-4V and Gemini 1.5 Flash across seven benchmarks.
• Dolma Open dataset from a diverse mix of web content, academic publications, code, books, and encyclopedic materials used to pretrain Olmo models; latest Dolma 3 supports standard pool and long-context mix.

Quantifiable outcome

• DR Tulu-8B costs approximately $0.00008 per query versus $1.80 for OpenAI Deep Research (about 22,500x cheaper)
• +7 more outcomes

Named customers14 records

Segments7 records

Ideal customer profiles6 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration1 record

AI capability16 records

Feature13 records

Partnerships

15 partnerships are on record, tiered core, moderate, notable, major and flagship.

• Cirrascale Cloud ServicescoreTechnology or Integration · 7 May 2026Manages the OMAI computing cluster deployed outside Austin on Nvidia Blackwell Ultra chips. The cluster is the first milestone in the $152M project to support AI development in materials science, biology, and energy.
• Stanford University Cong LabmoderateStrategic or Co-development Partner · 5 May 2026Piloted MolmoAct 2 for wetlab operations involving CRISPR gene editing. Researchers found the model showed strong potential for assisting with precision benchtop work, helping validate MolmoAct 2 for two-armed robotic tasks in real scientific settings.
• Carnegie Mellon UniversitymoderateStrategic or Co-development Partner · 26 March 2026Joint research on persona effects in AI chatbots; tests on six large language models with over 2,400 prompts found that expert personas increase errors by making outputs more confident and authoritative but less reliable.
• MicrosoftnotableOthers · 23 March 2026Microsoft hired former Ai2 CEO Ali Farhadi and core OLMo researchers to its Superintelligence team led by Mustafa Suleyman, reducing dependence on OpenAI. Ali Farhadi was hired as a corporate vice president. The departures reflect the broader competitive dynamic where commercial frontier labs can outbid non-profits on talent.
• AI2 Incubator ($80M fund)majorOthers · 7 October 2025Ai2's affiliated incubator program entering its next phase with an $80M fund to support AI startups built on Ai2's open research. Portfolio includes spinouts such as Vercept (founded by former AI2 research leaders, raised $16M seed led by Fifty Years).
• National AI Research Resource (NAIRR) PilotmoderateStrategic or Co-development Partner · 25 August 2025Ai2 participates in the NAIRR pilot launched by the NSF and 10 federal agencies with 25 private/nonprofit partners including NVIDIA, Google, Microsoft, Meta, AWS, and OpenAI. The pilot democratizes AI research access by providing advanced computing, datasets, models, and training to U.S.-based researchers.
• U.S. National Science Foundation (NSF)flagshipStrategic or Co-development Partner · 22 August 2025NSF contributed $75 million to the $152 million Open Multimodal AI Infrastructure to Accelerate Science (OMAI) project led by Ai2. The project creates a national-level, fully open AI ecosystem to drive scientific discovery and advance the science of AI itself, aligning with the White House AI Action Plan.
• NVIDIAflagshipStrategic or Co-development Partner · 22 August 2025NVIDIA contributed $77 million (in addition to NSF's $75M) to the $152M OMAI project. Provides Blackwell Ultra chips for Ai2's computing cluster and brings in advanced ML approaches and top research talent. NVIDIA also partnered with NSF on the broader National AI Research Resource (NAIRR) pilot.
• University of Pennsylvania (Penn Engineering)moderateStrategic or Co-development Partner · 21 July 2025Collaboration on CoSyn, a tool using open-source AI models to generate synthetic scientific figures, charts, and tables. CoSyn-trained models matched or outperformed GPT-4V and Gemini 1.5 Flash across seven benchmark tests. Full CoSyn code and dataset released publicly.
• University of Washington (Paul G. Allen School of Computer Science & Engineering)flagshipStrategic or Co-development PartnerLong-term foundational research partnership across AI model development, with co-located researchers and joint publications. UW's Allen School is named after Paul Allen, who founded both institutions. Joint work on open models (Olmo, Molmo, DR Tulu) and shared scientific infrastructure. UW board members (Ana Mari Cauce, Ed Lazowska) sit on Ai2's board.
• Google (Google Cloud)majorTechnology or IntegrationStrategic partnership supporting development of Ai2's open AI models and AI platforms that accelerate scientific research and perform modeling for environmental challenges. Google Cloud provides infrastructure for model training and deployment, including OlmoEarth wildfire risk work with NASA JPL and Google.org.
• IUCN (International Union for Conservation of Nature)majorStrategic or Co-development PartnerPartnership to equip governmental and non-governmental organizations with advanced AI to safeguard the high seas following the signing of the Biodiversity Beyond National Jurisdiction Agreement (BBNJ) at the UN SDG Summit. The Skylight platform provides AI technology at no cost to fast-track implementation of the UN High Seas Treaty.
• NASA Jet Propulsion Laboratory (JPL)majorStrategic or Co-development PartnerWorking with NASA and Google.org, OlmoEarth enables rapid model development for next-generation wildfire risk systems. Kimberley Miner (NASA JPL) cited that the ability to easily fine-tune and deploy models is accelerating their fire risk work.
• NOAA Geophysical Fluid Dynamics Laboratory (GFDL)majorStrategic or Co-development PartnerPartnership to design new global storm-resolving model (GSRM) simulations with less than 5 km horizontal spacing, and use them to train ML climate emulators (ACE, SamudrACE, FloeNet). Joint development of next-generation climate modeling infrastructure.
• Lawrence Livermore National Laboratory (LLNL)majorStrategic or Co-development PartnerDepartment of Energy-funded partnership to design GSRM simulations and use them for ML climate emulator training. Joint work on FloeNet (sea ice emulator), SamudrACE (coupled climate modeling), and HiRO-ACE (storm-resolving model emulator).

Scale indicators19 records

Recent moves7 records

Expansion highlights6 records

Direct peers

• Hugging Face: Primary distribution platform for open AI models, datasets, and benchmarks. Hugging Face is the de facto hub where Ai2's allenai organization publishes Olmo, Molmo, Tulu, and evaluation suites, and where Ai2's open-first community-led GTM motion is centered.
• Mistral AI: European frontier AI lab releasing open-weight foundation models (Mistral, Mixtral) competing in the same open-weight LLM category as Ai2's Olmo and Tulu families. Both target developers and enterprises that prefer open-weight deployments.
• Stability AI: Open AI research and product company releasing open foundation models (Stable Diffusion, Stable LM) under permissive licenses. Shares Ai2's open-first community-led GTM and developer-focused distribution, though Stability is commercial and image-focused.
• EleutherAI: Non-profit AI research lab pioneering open-source large language models (GPT-Neo, GPT-J, The Pile). Closest direct peer in mission, structure (non-profit), and open-first research approach; smaller in scale than Ai2.
• LAION: Non-profit open-source AI research organization that produced the LAION-5B dataset powering Stable Diffusion. Shares Ai2's non-profit, open-data, open-model ethos and developer-community distribution model, though focused on image datasets.

Broad incumbents

• Meta AI (FAIR): Meta's Fundamental AI Research lab is the highest-profile open-weight LLM producer (LLaMA family). Operates as a much larger incumbent in the same open-weight frontier-model category as Ai2's Olmo and Molmo, with vastly more compute and engineering resources.
• OpenAI: Frontier AI lab producing closed-weight commercial models (GPT-4o, GPT-5, OpenAI Deep Research). Functions as the contrasting incumbent to Ai2's open-first approach, and the price benchmark Ai2's DR Tulu is positioned against (22,500x cost differential).
• Anthropic: Frontier AI safety lab producing Claude family models. Competes for AI safety institute mindshare (UK AISI uses both AstaBench and Claude) and for the broader scientific/enterprise research market Ai2 targets.
• Google DeepMind: Frontier AI research lab behind Gemini, AlphaFold, and major weather/climate AI work. Overlaps with Ai2 in scientific discovery (Gemini 3 Pro is a DR Tulu benchmark competitor) and in climate modeling partnerships (NASA, NOAA ecosystems).

Emerging players

• Together AI: Open-model AI cloud platform hosting and serving open-weight models (including many of Ai2's via Hugging Face integrations). Comparable developer-first API distribution layer; commercial where Ai2 is non-commercial.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks6 records

Key highlights7 records

Customer concentration

• GitHubgithub.com/allenai
• LinkedInlinkedin.com/company/allen-institute-for-artificial-intelligence
• Xx.com/allen_ai
• Facebookfacebook.com/alleninstituteforai

Revenue estimate

Valuation estimate

Number of profiles

Profiles9 records

Funding overview

Funding rounds1 record

Investors2 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments16 records

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Allen Institute for Artificial Intelligence