Shanghai Artificial Intelligence Laboratory Open-Sources Weather Model ATRIA Dawn Preview

On September 14, Shanghai Artificial Intelligence Laboratory released the ATRIA Dawn Preview model as open source. On the same day, Muxi Co., Ltd. said it had completed Day0 adaptation for the model using its self-developed general-purpose GPU and MXMACA software stack, achieving availability immediately upon release.

For weather forecasting, ATRIA Dawn Preview is based on more than 100GB of global weather data, with a ViT backbone of over 0.4 billion parameters. After 45k training steps, it learns how 69 weather variables evolve over time and can forecast global weather one week ahead within one minute. While using less training data, the model outperforms the classic forecasting model FourCastNet on some metrics.

Since December 2025, Muxi has completed Day0 adaptation for 37 mainstream flagship models, covering developers including Zhipu, Alibaba Qwen, MiniMax, DeepSeek, StepFun and Tencent Hunyuan. The MXMACA software stack currently supports more than 40 AI frameworks and over 1,000 models, has been tested against more than 6,000 open-source projects, and fully supports 2,410 GPU operators in PyTorch 2.8.