allenai-org

MolmoWebAn open agent for automating web tasksLearn moreAI for the planet - Powerful platforms for planetary problemsOur AI research turns planetary data into action—deeper climate science, stronger agriculture and food security, proactive wildfire management, and lasting protection for wildlife and ecosystems.Explore our projectsOpen models - Olmo: the truly open LLMOur open-first approach empowers researchers and developers to advance the science of language models and use them in new and exciting ways.Dive into OlmoAI for science - An agentic ecosystem that advances scientific discoveryAsta unites three pillars for scientific AI: Asta agents that aid researchers with complex tasks, AstaBench sets rigorous benchmarks for any agent, and Asta resources provides tools and standards for building and testing.Meet AstaEmbodied AI - Bringing 3D reasoning into the openWe lead cutting-edge research to develop the next generation of intelligent robots, safely trained in advanced simulation environments to automate routine tasks and improve daily life.Learn moreAI news - Stay up to date with the latestLearn about our latest releases and go behind the scenes in our foundational research.Start readingResearch principles - Research is in our DNAOur research is guided by a strong set of open, collaborative, and inclusive principles that are designed to create safe, ethical, and trustworthy AI systems and resources.Dive deeperOur partnersDeep collaboration is the only way to build advanced, open AI. We partner closely with a wide range of external organizations to deliver breakthrough moments.University of WashingtonWe partner with the UW's Paul G. Allen School of Computer Science & Engineering on a wide variety of our foundational and applied research, amplifying our impact and supporting the next generation of AI researchers and developers.Visit the Allen SchoolNational Science FoundationOur partnership with the NSF supports the Open Multimodal AI Infrastructure to Accelerate Science project – the creation a national level, fully open AI ecosystem to drive scientific discovery through AI, while also advancing the science of AI itself.Learn more on our blogGoogleOur partnership with Google Cloud supports development of our state-of-the-art, truly open AI models, and our AI platforms that accelerate scientific research and perform modeling for a variety of evironmental challenges.More about our partnershipCareers - Join usInternship opportunities, cutting-edge research programs, and exciting new jobs. There are a whole host of ways for you to get involved at Ai2.Current jobsView open rolesInternshipsExplore internshipsSubscribe to receive monthly
updates about the latest Ai2 news.First NameLast NameEmailSign up --- About us - We’re Ai2We are a Seattle based non-profit AI research institute founded in 2014 by the late Paul Allen. We develop foundational AI research and innovation to deliver real-world impact through large-scale open models, data, robotics, conservation, and beyond.Our mission - Building breakthrough
AI to solve the world’s
biggest problems.Our people behind the AIWe unite the best and brightest scientific and engineering minds to explore the potential of truly open AI. Through close collaboration, we rapidly identify, define, and act on the most exciting and promising new ideas in our industry. To us, a variety of skills, experiences, and backgrounds is key to building the safest, most effective open AI technology. We invest in research into accessibility and inclusive user experiences, intentional hiring practices, and community building.See our facesOur core values - What we care aboutOpennessTrue openness means more than open source. By keeping an open mind and sharing everything we make, we bring the AI community together to answer research questions and build solutions that advance the whole field together.ScienceWe’re grounded in science. Scientific methods help us find breakthroughs that are rigorous and reproducible and allow us to deliver innovations at the forefront of AI.ImpactFocus lets us take bigger bets and optimize the impact of every endeavor to prioritize what matters most. This means we concentrate on the projects that will make the biggest difference.CollaborationGreatness is never achieved alone. We foster the conditions for deep collaboration internally and work closely with partners externally to stay at the forefront of impactful AI innovations.Photo by Kevin Cruff, © Estate of Paul G. AllenThe person who started it all - Paul AllenPhilanthropist and Microsoft co-founder Paul Allen founded Ai2 in 2014 to find transformative ways to develop AI to address some of the world’s biggest challenges. It’s his vision that enables us to push the boundaries of what’s possible.Learn about PaulMeet the boardEach one of them brings unique expertise and deep experience to help guide Ai2’s direction and deliver on our mission.Jody AllenTrustee, Paul G. Allen Trust, Co-founder and Chair, Allen Institute, Co-founder and Chair, Allen Family Philanthropies Ana Mari Cauce33rd President of the University of Washington Steve HallVenture Partner, Cercano Management Bill Hilf (Chairman)Ed LazowskaProfessor, Bill & Melinda Gates Chair Emeritus, UW Paul G. Allen School of Computer Science & Engineering Scientific excellenceOur scientific advisory board members offer invaluable guidance in our pursuit of impactful research in AI.Adam CheyerVP of AI Experience at Airbnb, co-founder of Siri David ForsythFulton Watson Copp Chair in Computer Science at the University of Illinois at Urbana-Champaign, Fellow of IEEE and ACM Eric HorvitzTechnical Fellow and Chief Scientific Officer of Microsoft, Fellow of AAAI, ACM, NAE, AAAS; AAAI President (2007-09) Mirella LapataProfessor at University of Edinburgh School of Informatics, BCS Karen Sparck Jones Award 2009, SIGDAT President 2018, Fellow of the Royal Society of Edinburgh, ACL, and Academia Europaea Tom MitchellFounders University Professor at Carnegie Mellon University, Fellow of AAAI and AAAS, AAAI Distinguished Service Award 2007 Aude OlivaDirector of Strategic Industry Engagement at the MIT Schwarzman College of Computing, Director of MIT-IBM Watson AI Lab, Senior Research Scientist at MIT Dan RothEduardo D. Glandt Distinguished Professor at the University of Pennsylvania and Chief AI Scientist at Oracle, Fellow of AAAS, ACM, AAAI, and ACL, John McCarthy Award (IJCAI) 2017 Subscribe to receive monthly updates about the latest Ai2 news.First NameLast NameEmailSign up --- OlmoOur fully open language model and complete model flow.Chat with OlmoBuild with OlmoThe Olmo 3 model familyPick a variant to explore weights, code and reports. Every card includes instant links to artifacts. Read the technical report32B-BaseAchieves strong results in programming, reading comprehension, and math problem solving, maintains performance at extended context lengths, and works well with RL setups.32B-ThinkCapable of reasoning through complex problems step by step. A strong platform for RL research and other advanced experiments that need serious horsepower.32B-InstructOur most capable fully open chat model to date. An instruction-tuned model built for chat, tool use, and multi-turn dialogue.7B-BaseA smaller, lighter-weight base model able to run on a wider range of hardware while delivering competitive performance.7B-ThinkDelivers strong reasoning capabilities at 7B scale, surfacing intermediate thinking steps for complex prompts at high efficiency.7B-InstructModel for efficient inference that handles multi-turn chat, tool use, and more.A complete model flowTo truly advance open AI development and research, the entire model flow – not just its endpoint – should be accessible and customizable. The model flow is the full lifecycle of an LM, starting with the data.Explore the Model FlowClick on any stage to learn more about it and download artifacts.Pretraining dataThe fully open mixture used to train Olmo from scratch—curated web, code, books, and scientific text—deduplicated and quality-filtered.Standard PoolLong Context MixMid-training dataTargeted continuation sets used to refine the base model mid-course. Higher-quality, domain-focused mixtures.DownloadPost-training dataCorpora used after pretraining for instruction tuning and preference-based optimization where applicable—supervised responses and comparison data.DownloadOpen-source toolsThese are the tools we use to make Olmo.OlmoCoreOur training framework for fast, easy configurationAccess OlmoCoreData preprocessing toolsDuplodocusUltra-efficient fuzzy de-duplicationDatamap-rsFor large-scale data cleaningOpen InstructOur post-training pipelineAccess Open InstructModel evaluationOLMESUtility for reproducible evalsDeconHelps remove test sets from training dataMore resourcesOlmoTraceTrace Olmo’s output back to the training dataFlexOlmoA new paradigm for language model training and data collaborationDocumentationGet started building your own projectBlogRead about our latest news and releasesWhat people are sayingUnlike current open LLMs, which limit access to their training data, architectures, or evaluation methodologies, Olmo stands out by providing full access.Janakiram MSVForbes“We see Olmo’s architecture as pushing the frontier of open-source model design... Open source is how we drive progress forward!Simon MoProject Co-Lead, vLLMOlmo is becoming the instrument through which the community builds the next layer of open, foundational intelligence.Anastasios AngelopoulosCEO, LMArenaOlmo has led to a wellspring of research and innovation that couldn't have been accomplished without a fully open model... Olmo ensures complete transparency and sets a strong foundation for transformative work.Clem DelangueCo-Founder & CEO, Hugging FaceAi2’s idea is to give the AI community full visibility into a state-of-the-art large language model in order to confront the problems with existing LLMsMark SullivanFast CompanyTransparency and performance are essential for developers to scale AI with open, U.S.-built models like Olmo 3Kari BriskiNVIDIABuilt for research— already making impactFrom unlearning to clinical NLP, Olmo is powering discoveries across domains. Explore how researchers are using fully-open models.Machine unlearning with Olmo-7BResearchers used Olmo-7B as a testbed for developing machine unlearning methods—removing specific data influence without retraining from scratch.See projectClinical NLP applicationsHealthcare teams leveraged Olmo checkpoints to explore clinical text analysis while preserving transparency around data and methods.See projectUnderstanding how LLMs learnOlmo’s openness—datasets, logs, and checkpoints—enabled fundamental studies into learning dynamics and scaling behaviors.See projectDeep dive with Olmo lead researchers Hanna Hajishirzi and Noah Smith on how - and why - we built Olmo 3, and what comes next.Subscribe to receive monthly updates about the latest Ai2 news.First NameLast NameEmailSign upWe must use local storage to remember your permissions. Can we also use cookies and external services according to our privacy policy to improve the browsing experience?Manage OptionsReject AllApprove All --- Tülu3Try Tülu 3Tülu 3 on the Ai2 blogTülu 3 is a leading instruction following model family, offering fully open-source data, code, and recipes designed to serve as a comprehensive guide for modern post-training techniques.ModelsExplore the collection of open-sourced instruct models created from our open data and recipes.Tülu 3 model familyDataThe underlying training data for fine-tuning processes is the most important piece of the puzzle but often the element with the least transparency. Tülu 3 changes that.Get the dataTrainingWe open source our scalable codebase for supervised finetuning (SFT), Direct Preference Optimization (DPO), Reinforcement Learning with Verifiable Rewards (RLVR), and all the other algorithms we considered when training Tülu.Training codeEvaluationWe're sharing the code base used to produce Tülu 3's results to make these evaluations more standardized and reproducible.Evaluation suiteDecontaminationPaperCheck out the Tülu 3 paper for more insights into the premise and the creation of the Tülu 3 collection.Read the paperBlogsTülu 3 represents the next era in open post-training. Check out our blog for more on this important new release from Ai2.Release announcementTechnical postOur philosophyEarly work in language model post-training followed a standard recipe pioneered by models like InstructGPT, consisting of instruction-tuning followed by preference fine-tuning. Since then, the sophistication and complexity of post-training approaches have continued to increase, however most successful post-training models offer limited information about their training data, code, or recipes. Tülu 3 pushes the boundaries of research in post-training and closes the gap between open and closed fine-tuning recipes. By openly sharing our data, recipes, and findings, we hope to uncover which paths for the open-source community will lead to success and which do not, enabling the community to explore new and innovative post-training approaches.Our approachThe Tülu 3 effort began with identifying key desirable capabilities for generalist language models, including knowledge, reasoning, mathematics, coding, instruction following, general chat, and safety – areas where current open post-training recipes often fall behind. Our success is rooted in careful data curation, rigorous experimentation, innovative methodologies, and improved training infrastructure. In particular, we produce Tülu 3 models through a four-stage post-training recipe on top of pre-trained language models (namely Llama 3 Base). This includes (1) careful prompt curation and synthesis, (2) supervised finetuning on our carefully selected mix of prompts and their completions targeting core skills, (3) combining both off- and on-policy preference data to apply preference tuning, and (4) a new RL-based method to enhance specific skills with verifiable rewards.The stages of development of Tülu 3's datasets, training methods, and evaluation suite.Our resultsTülu 3 models achieve state-of-the-art performance across our multi-skill evaluation compared to models of an equivalent size and some closed API-based models.Subscribe to receive monthly updates about the latest Ai2 news.First NameLast NameEmailSign upWe must use local storage to remember your permissions. Can we also use cookies and external services according to our privacy policy to improve the browsing experience?Manage OptionsReject AllApprove All