The Allen Institute for Artificial Intelligence, commonly known as AI2, is a nonprofit Artificial intelligence research institute based in Seattle, Washington. It was founded in 2014 by Microsoft co-founder Paul Allen, who had previously funded the Allen Institute for Brain Science and sought to apply similar philanthropic ambition to AI research. Oren Etzioni served as the institute's first chief executive from its founding until 2022.
History
AI2 positioned itself from the outset as a research-first counterweight to the commercial AI labs emerging in Silicon Valley. Early projects included Semantic Scholar, a free academic search engine that applies Machine learning and Natural language processing to scientific literature, and AllenNLP, an open-source library built on PyTorch for NLP research. The institute also incubated PRIOR, a computer-vision group, and supported spinout companies that grew out of internal research groups. Paul Allen's death in 2018 did not end the institute; it continued under his estate's endowment and later leadership including Ali Farhadi.
OLMo and open research
AI2's most visible contribution to the broader Large language model landscape came with OLMo (Open Language Model), released starting in 2024. Unlike most frontier models from labs such as OpenAI, Anthropic, and Google DeepMind, OLMo was released with its full training data (the Dolma corpus), training and evaluation code, and intermediate checkpoints, not just final weights. This made it one of the few genuinely Open-weights models models where researchers could study the entire training process rather than only the finished artifact, a distinction AI2 argued mattered for reproducible science and for studying emergent capabilities and training data effects. Subsequent OLMo releases improved benchmark performance while maintaining the fully-open release philosophy, and AI2 published companion research on data curation, fine-tuning, and instruction tuning.
Reception and impact
AI2's fully open approach has been cited by researchers and policymakers as a reference point in debates over AI governance and the transparency of training data used by commercial labs, several of which face copyright lawsuits over undisclosed training corpora. Because OLMo models generally trail top proprietary and even top open-weight models like Llama or Qwen on many benchmarks, AI2's contribution is often framed less as a competitive claim to state-of-the-art capability and more as an infrastructure and transparency project for the wider research community. The institute continues to operate as a nonprofit alongside for-profit AI labs, funding open tooling, datasets, and models that downstream researchers and smaller companies can build on without licensing restrictions.