Ai2
Why it matters
In a field where the leading models are closed, and where even "open" models usually share only their weights, Ai2 releases everything: weights, training data, training code, evaluation code and intermediate checkpoints. That lets outside researchers study how training data shapes behavior, reproduce results and build on the work without guessing, which is otherwise impossible for models of this scale.
It also builds free tools that scientists use every day. Semantic Scholar, its AI-powered academic search engine, indexes more than 200 million papers and serves millions of researchers, and its open data feeds many other research tools.
Deep Dive
The Allen Institute for AI, styled Ai2, is a nonprofit research lab in Seattle. Microsoft co-founder Paul Allen created it in 2014 with a mission of "AI for the common good". Its founding CEO was computer scientist Oren Etzioni. Ali Farhadi, a computer vision researcher and one of the co-authors of YOLO, led it from 2023 until March 2026; Peter Clark has been interim CEO since.
Ai2 has a few hundred researchers and engineers, tiny next to the big labs, but its influence on open AI research is far larger than its size suggests.
Language Research Roots
Ai2 was a major force in NLP before the LLM era. Its researchers created ELMo (2018), the contextual word embeddings that directly preceded BERT and helped start the pre-training revolution, and AllenNLP, a widely used research library. Its Aristo project built systems to answer science exam questions, a test of reasoning long before today's leaderboards, and its teams created many well-known datasets and benchmarks.
Open Models
- OLMo (from 2024): a family of fully open language models, released with their training data (the Dolma corpus), code, logs and checkpoints. Later versions, OLMo 2 and Olmo 3, became competitive with other open models of similar size while keeping every step of development public.
- OLMoTrace: a tool that traces a model's output back to matching passages in its training data in real time, which is possible only because the training data is public.
- Tülu: open recipes, data and code for post-training (instruction tuning, preference tuning and reinforcement learning), the stage that labs usually keep secret.
- Molmo (2024): open multimodal models that can point at things in an image, useful for counting, grounding and controlling user interfaces. They were trained on a dataset collected by having annotators describe images aloud.
- olmOCR (2025): an open toolkit and model for converting PDFs and scanned documents into clean text at very low cost.
AI for Science
Beyond Semantic Scholar, Ai2 develops AI agents to assist with scientific work, including Asta, a platform for literature review, data analysis and benchmarking of scientific agents. In 2025 the US National Science Foundation and NVIDIA announced a combined $152 million for an Ai2-led project to build open multimodal models for scientific research.
It also runs applied programs for the environment, such as Skylight, which detects illegal fishing from satellite data, and EarthRanger, which is used to protect wildlife.
Embodied AI and Vision
Ai2's vision and robotics teams built AI2-THOR, a simulated indoor environment widely used to train and test embodied agents, and Objaverse, a huge open dataset of 3D objects. The original YOLO work also has ties to the institute through Farhadi's group at the University of Washington. An incubator at Ai2 has spun out a number of startups, among them XNOR.ai, an edge AI company that Apple acquired in 2020.
Open vs. Open
Ai2's work sharpens an important distinction. "Open weights" means you can download and run a model. "Open source" in the fuller sense means you can also see what it was trained on and how. Only the second allows real scientific scrutiny: checking benchmarks for contamination, studying memorization, understanding where a behavior comes from. Few organizations can afford to work this way, and Ai2 is the most prominent one that does.