Ai2

nonprofit

usa · Founded 2014

The Allen Institute for AI (Ai2) is a nonprofit research institute founded in 2014 by Microsoft co-founder Paul Allen, headquartered in Seattle. Funded by the Paul Allen estate ($3.1B foundation) with annual budget exceeding $100M, plus $152M from NSF and NVIDIA for the OMAI scientific-AI initiative (2025). ~300 researchers.

Ai2's defining contribution is the OLMo (Open Language Model) ecosystem — the first competitive LLM to release everything: weights, training data, code, logs, and thousands of intermediate checkpoints under Apache 2.0. OLMo 1 (Feb 2024) established the paradigm; OLMo 2 (Dec 2024, up to 32B/6T tokens) was the first fully open model to outperform GPT-3.5 Turbo; OLMo 3 (Dec 2025, 32B, 65K context, Think variants) introduced 3-stage training with mid-training and long-context phases. The latest OLMo Hybrid (Mar 2026) combines Gated DeltaNet + Transformer at 3:1 ratio, reaching OLMo 3 quality in 35% fewer tokens.

The OLMo stack includes Dolma (3T+ token open pretraining corpus), Tülu (post-training framework introducing RLVR — Reinforcement Learning with Verifiable Rewards), OLMoE (first fully open MoE LLM, 6.9B/1.3B active, 64 experts), and OLMES (evaluation standard). In March 2026, Ai2 lost CEO Ali Farhadi and several key researchers to Microsoft's Superintelligence team, though all 2026 programs remain funded.

Website Wikipedia HuggingFace GitHub OpenRouter
nonprofitopen-sourceopen-weightreasoning

People

News

Outputs (20)

OlmPool: Cracks in the Foundation

paper

Studies how seemingly minor architectural choices impact long-context extension in language models. Provides a pool of 7–8B parameter models tra…

BAR: Branch-Adapt-Route

paper

Modular post-training recipe that trains independent domain experts separately then merges them into a Mixture-of-Experts model. Each expert is a two-…

OLMo Hybrid

model

First fully open hybrid Transformer + linear RNN model at scale. Gated DeltaNet + Transformer attention at 3:1 ratio (3 DeltaNet layers per 1 attentio…

OLMix

paper

Framework for data mixing across the full LM development lifecycle. Solves the "evolving domain problem" where domains are added, removed, or revised…

Molmo 2

model

Open VLM family expanding Molmo's image strengths to video and multi-image understanding. 4B (Qwen 3), 8B (Qwen 3), and 7B-O (OLMo backbone, fully ope…

OLMo 3

model

7B and 32B dense Transformers with 65K native context (up from 4K). 3-stage training: main pretraining (~95%), mid-training (100-200B tokens of code/m…

FlexOlmo

model

New paradigm for LLM training via data collaboration. Data owners independently train MoE expert modules on their private data, then contribute expert…

IFBench

eval

58 diverse, programmatically verifiable output constraints for evaluating instruction-following generalization. Core finding: models strongly overfit…

DataDecide

paper

Most extensive open suite of controlled pretraining experiments over data and scale: 25 corpora (Dolma, DCLM, RefinedWeb, C4, FineWeb), sizes up to 1B…

WebOrganizer

paper

Framework for constructing meaningful domain taxonomies over web corpora. Introduces topic and format taxonomies (24 categories each), distilling Llam…

OLMo 2

model

"2 OLMo 2 Furious." Dense Transformers at 7B (4T tokens), 13B (5T), and 32B (6T tokens, 1.5 epochs). 4K context. Introduces Dolmino Mix late-stage cur…

Language Model Ladders

paper

Develops task-specific scaling laws using small "ladder" models (costing 1% of target compute) to predict large model performance within 2 percentage…

Tülu 3

model

Post-training framework introducing RLVR (Reinforcement Learning with Verifiable Rewards). Applied to Llama 3.1 at 8B, 70B, and 405B. Surpasses instru…

Molmo

model

Open multimodal VLM family with fully open training data (PixMo). Pioneered image pointing capabilities. Trained in two stages: dense captioning pre-t…

OLMoE

model

First fully open MoE LLM: weights, training data (OLMoE-Mix), code, and logs. 6.9B total / 1.3B active per token with 64 experts per layer (top-8 drop…

SciRIFF

dataset

137K instruction-following demonstrations across 54 scientific literature tasks (extraction, summarization, QA, claim verification, classification). S…

OLMo

model

First competitive LLM to release everything: weights, training data (Dolma), training code, logs, and 500+ intermediate checkpoints. 1B and 7B dense T…

Dolma

dataset

Largest fully open pretraining corpus at release: ~3T tokens from Common Crawl (2,180B), GitHub code (342B), Reddit (80B), Semantic Scholar/peS2o (57B…

Paloma

dataset

Perplexity-based benchmark measuring language model fit across 585 textual domains from 18 sources, rather than assuming perplexity on one distributio…

GPQA: Graduate-Level Google-Proof Q&A

eval

198 "Google-Proof" graduate-level science questions (the Diamond subset) in biology, physics, and chemistry. Designed so that even skilled non-experts…