Allen Institute for AI

The Allen Institute for AI (Ai2) is a nonprofit artificial intelligence research institute based in Seattle, Washington, founded in 2014 by Microsoft co-founder Paul Allen. It conducts fundamental and applied AI research for the common good and is best known for releasing OLMo (Open Language Model), a family of state-of-the-art large language models released together with their training data, code, and checkpoints ^[1][2].

History

Ai2 was established by Paul Allen, who had previously founded the Allen Institute for Brain Science (2003) and the Allen Institute for Cell Science (2014), with a mission to use AI to solve major scientific and societal problems. Oren Etzioni served as its first CEO, and the institute built early renown through projects such as Semantic Scholar, an AI-powered academic search engine ^[1].

In May 2023, Ai2 announced OLMo, an open language model project, and in February 2024 it released the 1B and 7B parameter variants along with the complete Dolma training dataset (a three-trillion-token English corpus) and all training code, intermediate checkpoints, and logs — making it the most open state-of-the-art model at the time ^[1][2]. The family expanded with OLMoE (September 2024), a mixture-of-experts model developed with Contextual AI, and OLMo 2 (November 2024) in 7B and 13B sizes, followed by a 32B variant in March 2025 that Ai2 said was the first fully open model to outperform GPT-3.5-Turbo and GPT-4o mini ^[1]. In November 2025 the institute released OLMo 3, which it said outperformed fully open models such as Stanford's Marin and commercial open-weight models like Meta's Llama 3.1, with support for contexts up to 65,000 tokens ^[3].

Key Concepts

The defining principle of the OLMo project is that "open" means the complete scientific record of a model: weights, training data, training code, evaluation code, and logs are all released, allowing researchers to reproduce, audit, and study how the model was built ^[2]. This contrasts with "open-weight" releases from commercial labs, which typically publish weights but not data or training details ^[1][2].

OLMo models are built on the transformer architecture and trained with conventional large-scale pretraining pipelines, while OLMoE applies mixture-of-experts sparsity (about 7 billion total parameters with roughly 1 billion active per token) for efficiency ^[1]. Ai2's goal is to keep AI an open scientific field — enabling researchers to study capabilities, biases, and safety properties from the inside — rather than a set of closed commercial products ^[2].

Applications

OLMo models are used by researchers studying interpretability, bias, and training dynamics; by universities and startups that need fully auditable models; and by organizations that want to fine-tune models without licensing restrictions ^[1][2][3]. The institute also develops open tools and benchmarks used across the field. In 2025–2026 Ai2 invested in a major computing cluster in a $152 million project backed by NVIDIA and the US National Science Foundation, expanding its capacity to pretrain large open models ^[3].

>See Also

• Open Weights
• Llama
• DeepSeek
• Mixture of Experts
• ILMU (Malaysian LLM)
• Sea-Lion (Malaysian LLM)