Ai2
57 Intel
The Allen Institute for AI (Ai2) is a nonprofit AI research institute founded in Seattle in 2014 by Microsoft co-founder Paul Allen. Its research spans deep learning, computer vision, natural language processing, and common-sense reasoning, and it develops open language and multimodal model families including OLMo, Tulu, and Molmo, publishing their weights, training data, and code under permissive licenses. It is led by interim CEO Peter Clark, a founding member and senior research leader.
Is this you? Sign in with X to claim this profile.
Intel
The NSF OMAI compute is live. Blackwell Ultra systems are now operational at Ai2 under a mandate that every training run be reusable and extensible, which is a different bet on GPU-hour return than the closed labs are making.
Ai2 open-sources AstaBrief 8B, a small model that turns a research question and retrieved literature excerpts into a cited report. It aims to match proprietary report quality with faster generation and cheaper serving, and ships as Fast mode in Asta.
Ai2 opened the rest of MolmoAct 2: fine-tuning scripts, the complete training datasets, all evaluation rollouts, and the recipe for its open tokenizer, with inference and training now integrated into Hugging Face's LeRobot so existing setups can take the model without retooling.
Ai2's byte-level retrofitting work behind Bolmo is now in Nature. New Hugging Face checkpoints extend the approach to Qwen and Llama, and Stage 1 checkpoints with frozen original weights give researchers a faster start on byte-level architectures.
Ai2's open earth-observation model family gets a positional-encoding rebuild. Swapping absolute embeddings for RoPE removes artifacts partners were seeing in patch embeddings and lifts kNN/linear-probe results at every size, from Nano to Base. Weights and tech report are public.
Nextcloud bundles Olmo 3 7B in its llm2 local LLM backend, chosen for its highest Ethical AI rating (open training data and software, self-hostable) and because it runs without 16 to 32 GB of VRAM. Admins configure it centrally for Assistant and Mail.
Tracking explains motion that already happened; planning needs the next few seconds. MolmoMotion takes an RGB frame, points marked on an object, and an action description, and predicts those points' 3D trajectories forward. It is released with models, data, a benchmark, and code.
A replay-based eval that drops a model into real tutoring transcripts at the moment a human tutor had to choose between giving help and holding back, with a simulated student on the other side. Dataset, code and tech report are open. Headline result: models over-help by default.
Ai2's AutoDiscovery agent analyzes datasets, proposes hypotheses and ranks results by Bayesian surprise; a University of Washington materials science class used it to surface leads in battery-aging data, then verified each by hand.