Ai2

Ai2

Ai2, the Allen Institute for AI, is a non-profit research institute in Seattle that releases fully open AI models. Its best-known family, Olmo, comes with more than the weights: Ai2 also publishes the pretraining and post-training data, the training code, intermediate checkpoints and evaluation tools, so researchers can trace how a model learned what it knows. Ai2 also builds Molmo multimodal models, the Tülu post-training recipes, and Asta, a set of AI agents and benchmarks for scientific research.

Production credibility: Ai2 was founded in 2014 by Microsoft co-founder Paul Allen and partners with the University of Washington's Paul G. Allen School, the US National Science Foundation and Google Cloud. The current Olmo 3 family comes in 7B and 32B sizes, each as Base, Think and Instruct variants, and the models are released under Apache 2.0. Molmo 2 adds 4B and 8B vision-language models, with strong video understanding, plus a fully open 7B variant built on Olmo. OlmoTrace lets users trace a model's output back to the training data it came from. In October 2026 Ai2 open-sourced AstaBrief, an 8B model for cited scientific reports. Ai2's site says its AutoDiscovery agent helped researchers at Providence Swedish Cancer Institute find an immune signal in invasive lobular breast cancer. All models, data and tools are free to download.

Key Features

• Olmo 3 language models in 7B and 32B sizes, Base, Think and Instruct
• Open pretraining and post-training data, code and checkpoints
• OlmoTrace for tracing model outputs back to training data
• Molmo 2 multimodal models for images, video, pointing and tracking
• Asta research agents and the AstaBench evaluation suite
• Open-source training and evaluation tools such as OlmoCore and OLMES

Ideal Use Case

AI researchers and universities that need to study how a model was trained, not just run it, including work on data attribution, contamination and training dynamics. Companies that require a permissive Apache 2.0 licence and full data provenance for compliance can also build on Olmo, and scientists can use Asta agents for literature and discovery work.

How Ai2 differentiates

Meta's Llama, Mistral and DeepSeek publish model weights but not the full training data or pipeline. Ai2's Olmo is one of the few model families where everything is open, from the data mixture to the training code and checkpoints, which makes it the reference point for reproducible research. The trade-off is that its largest models are smaller than the biggest open-weight releases from commercial labs.

FAQ

Q: What is Ai2? A: Ai2 is the Allen Institute for AI, a non-profit research institute in Seattle founded in 2014 by Paul Allen. It builds fully open AI models such as Olmo and Molmo, plus scientific AI tools like Asta.

Q: What is Olmo? A: Olmo is Ai2's family of fully open language models. The current Olmo 3 family includes 7B and 32B models in Base, Think and Instruct variants, released with their training data, code and checkpoints.

Q: Are Ai2 models free to use? A: Yes. Ai2 releases its models, datasets and tools for free download. Olmo 3 models are published under the Apache 2.0 licence, which allows commercial use.

Q: What makes Olmo different from Llama? A: Llama releases model weights, while Olmo also releases the full training data, training code, intermediate checkpoints and evaluation tools, so the whole training process can be inspected and reproduced.

Q: What is Asta? A: Asta is Ai2's set of AI agents, benchmarks and tools for scientific research, including research assistants for literature tasks and the AstaBench evaluation suite.

tl;dr

Ai2, the Allen Institute for AI, is a non-profit that releases fully open models, including the Olmo 3 language models and Molmo 2 multimodal models, along with their training data and code. Everything is free, and Olmo 3 is licensed under Apache 2.0.