Atria Dawn: A New Model Trained for Scientific and Agent Tasks

Atria ASI released Atria Dawn Preview, a base model trained for scientific and complex agent tasks using a verifiable experience pipeline. It can predict global weather for the next week in one minute without internet. Two-thirds of the 96-member team are students.
Key insights
Training and Capabilities
Atria Dawn is trained using a verifiable experience pipeline: only effective experiences tested in real environments with metrics and evidence closure are incorporated into the model. It can learn from failures. In a completely offline environment, it can predict global weather for the next week within one minute.
Team Composition and Ecosystem
About two-thirds of the 96 main members are students (65 students) from Shanghai AI Lab, Fudan, CAS Software Institute, Renmin University, CAS Automation Institute, East China Normal, and HIT. The model comes with a dedicated agent harness (scaffold) for collaborative work, enabling a wide range of tasks.