AI2

AI2, the Allen Institute for AI, is a Seattle-based nonprofit AI research institute founded by Microsoft co-founder Paul Allen in 2014. Its work includes the Olmo open language models, Tülu post-training recipes, Molmo vision-language models and Semantic Scholar’s research-discovery tools. Across these projects, it releases training data, code and evaluation resources alongside models, giving researchers access to more of the development process than final weights alone.[3][1][4][5][6][13]

Before its open-model releases, Ai2 developed systems for finding scientific knowledge and answering science questions. Semantic Scholar launched at Ai2 in 2015 and remains a research and product-development team within the institute. Its in-house models process and classify papers, identify connections and surface information to help researchers discover and understand literature.[13][10][4]

Semantic Scholar reports indexing over 200 million academic papers and partnering with more than 50 publishers and scholarly societies on its undated about page. Papers enter the index through publisher partnerships, data providers and web crawls. Its open resources serve different purposes: the Semantic Scholar Academic Graph (S2AG) dataset and APIs distribute paper records, while the Semantic Scholar Open Research Corpus (S2ORC) supports NLP and text-mining research over scientific text.[13]

The Aristo project supplies earlier context for Ai2’s machine-reasoning research. In a paper first submitted in September 2019 and revised in February 2021, Peter Clark, Oren Etzioni and their coauthors reported scores above 90% on Grade 8 New York Regents Science Exam questions and above 83% on corresponding Grade 12 questions. Both results concerned non-diagram, multiple-choice questions on unseen tests; they excluded diagram and free-response questions and did not establish general question-answering mastery.[10][12]

In summer 2022, the AllenNLP team released AI2 Tango, a free tool for building machine-learning experiments from reusable, repeatable steps. That October, Aristo team director Peter Clark became interim CEO, starting on October 1 while Ai2 sought a permanent chief executive.[12]

In his January 2024 interview, research scientist Nathan Lambert described CEO Ali Farhadi’s effort to broaden Ai2’s emphasis from papers toward model releases and policy participation. Lambert said that direction helped draw him to the institute, where he worked on Tulu 2, an instruction-tuned model trained with Direct Preference Optimization (DPO).[1]

Lambert recalled working with graduate student Hamish Ivison to scale the Zephyr DPO recipe to a 70-billion-parameter model, using JAX-based code and TPUs available for research. He described the released model as a successful first run: “we did no ablations, didn't change any parameters.” His account illustrated an early Tulu 2 experiment; later Tülu development involved extensive data-mixture and hyperparameter experiments.[1][5]

Ai2 launched its first Open Language Model, OLMo, on February 1, 2024. The initial batch comprised four 7B-scale variants with different architectures, optimizers and training hardware, plus a 1B model, all trained on at least two trillion tokens. Collaborators included AMD, Finland’s CSC–IT Center for Science, Mosaic/Databricks, Harvard’s Kempner Institute and the University of Washington’s Allen School.[4]

OLMo’s framework provides Dolma training data and the code that produces it, model weights, training and inference code, logs, evaluation tools and more than 500 checkpoints per base model, saved every 1,000 training steps. These intermediate artifacts support experiments on data selection, training methods and instabilities. Code, weights and intermediate checkpoints are released under Apache 2.0. The OLMo announcement also describes adaptation through Open Instruct and Tulu 2 data: supervised fine-tuning produces the SFT checkpoint, followed by DPO for the Instruct version.[4]

Tülu 3, announced on November 21, 2024, developed the post-training side of this work using Meta’s Llama 3 Base models at 8B and 70B. Its recipe combines prompt curation and synthesis, supervised fine-tuning, DPO using responses from both Tülu and other models, reinforcement learning with verifiable rewards, and standardized evaluation. Ai2 released datasets, curation tools, decontamination scripts, training code and evaluation resources.[5]

Ai2 reported collecting 939,344 prompts for Tülu 3, with 57% drawn from public resources and 43% generated in-house. Its new synthetic supervised-training responses came from GPT-4o or, for coding, Claude 3.5 Sonnet. GPT-4o also judged candidate responses for helpfulness, instruction-following, honesty and truthfulness to construct preference pairs. After comparing preference algorithms including PPO, the team chose DPO for simplicity and experimental efficiency, reporting roughly similar results with better tuning and length normalization.[5]

Reinforcement Learning with Verifiable Rewards (RLVR) replaces a learned reward model with a verification function. For tasks such as mathematics and constrained instruction-following, answer matching or constraint checks supply a binary reward when an output is verifiably correct. Ai2 reported gains over Tülu 3’s DPO checkpoints on MATH, GSM8K and IFEval. Verifiability still depends on reward design: the July 2025 discussion with Lambert described how code models can game unit tests rather than implement correct logic, and how learning reliable tool use presents further difficulties.[5][2]

Ai2 introduced Molmo on September 25, 2024. The vision-language family answers questions about images and can point to visual elements, grounding answers in image locations. Its architecture combines multiscale image crops, a vision encoder, a connector that projects and pools visual tokens, and a decoder-only language model. Training proceeds through image-caption pretraining and supervised fine-tuning, updating all parameters without RLHF.[6]

Molmo’s PixMo data collection asked annotators to describe images aloud for 60–90 seconds, including spatial relationships and small details. Ai2 reported gathering spoken descriptions for 712,000 images across 70 topics, then transcribing and refining them with language models. The team said speech produced richer descriptions faster than typing, while retaining recordings as annotation receipts. This approach avoided distilling image captions from proprietary vision-language models, although it used language-only models for text processing.[6]

Molmo’s first release included four sets of weights: MolmoE-1B, Molmo-7B-O, Molmo-7B-D and Molmo-72B. PixMo datasets and training and evaluation code followed in a second release. Openness varied across pretrained components: all released models used OpenAI’s CLIP vision encoder, trained on closed data, while language backbones included Ai2’s OLMo and OLMoE and external Qwen2 models. Ai2’s human evaluation reported 325,231 pairwise comparisons across 27 models, collected from 870 annotators.[6]

Olmo 3, announced on November 20, 2025, expanded the release model into a documented “model flow”: data, code, weights and checkpoints spanning pretraining, targeted mid-training, long-context extension and post-training. The family included 7B and 32B Base and Think models, a 7B Instruct model for chat and tool use, and a separate 7B RL Zero route for reinforcement-learning experiments from base models. Think training followed supervised fine-tuning, DPO and RLVR.[7]

Ai2 assembled a roughly 9.3-trillion-token Dolma 3 corpus and selected a 5.9-trillion-token pretraining mix from it, with separate data mixtures for later training stages. It reported using up to 1,024 H100 GPUs for pretraining, 128 for mid-training and 256 for post-training. Moving supervised fine-tuning from Open Instruct to Olmo Core reportedly increased throughput eightfold; changes to batching, threading and weight updates made RL training four times more efficient.[7]

A later update labeled December 12 introduced Olmo 3.1 Think 32B and Instruct 32B, plus refreshed RL Zero code and math checkpoints. Ai2 said the Think model extended its earlier reinforcement-learning run by 21 days on 224 GPUs. The Instruct model applied the 7B instruction-training recipe at 32B scale for chat, tool use and multi-turn dialogue.[7]

On May 7, 2026, Ai2 reported that $152 million in combined NSF and NVIDIA project support, awarded in 2025, was becoming operational through the Open Multimodal AI Infrastructure for Science (NSF OMAI). The project combined an NSF cooperative agreement with a $77 million NVIDIA investment. Its B300 cluster was deployed and managed with Cirrascale Cloud Services to support training and experimentation across language, multimodal and scientific research.[9]

Ai2 leads NSF OMAI with co-principal investigators from the University of Hawai‘i Hilo, the University of New Hampshire, the University of New Mexico and the University of Washington. Noah A. Smith serves as principal investigator and Ai2 senior research director. The project’s emphasis is reusable open artifacts—data, checkpoints and methods that other teams can inspect and extend.[9]

In a May 2026 interview, interim CEO Peter Clark outlined priorities in model science and reliability, scientific discovery, embodied AI and environmental applications. Asta, an agentic ecosystem for research, included ScholarQA, AutoDiscovery and Theorizer. Clark also highlighted early embodied systems MolmoAct and MolmoBot and Earth-systems work through OlmoEarth.[8]

AutoDiscovery progressed from an open-ended scientific-discovery research system to a managed solution where researchers upload structured datasets, generate and test hypotheses, and inspect the code and statistical analysis behind results. On August 27, 2026, Ai2 announced a partnership with the Paul G. Allen Research Center (PARC) at Providence Swedish Cancer Institute to apply it across cancer-research datasets.[8][11]

The joint research team was led by Providence researchers Kelly Paulson and Sasha Stanton and Ai2 senior research scientist Bodhisattwa Majumder. Ai2 reported that applying AutoDiscovery to The Cancer Genome Atlas surfaced a stronger immune signature in invasive lobular breast cancer, followed by validation in an independent dataset and laboratory analysis of tumor samples. The finding suggested further immunotherapy research; it did not establish treatment efficacy. Providence was also standing up AutoDiscovery inside its own cloud environment, with PARC’s computational research team installing, operating and supporting the system so protected data would remain within Providence.[11]