SenseTime unveils Sage edge-side multimodal intelligent agent base model, leading a new era of automotive intelligent agents.

On April 22, 2026, SenseTime launched the Sage edge-side multimodal intelligent agent base model, adopting the MoE architecture with 32 bytes of parameters and 3 bytes of activation parameters, becoming the first base model to achieve complex intelligent agent capabilities on the vehicle side. Sage performed exceptionally well on PinchBench, achieving a task completion rate of 94%, surpassing many mainstream international cloud/edge large-scale models. Sage incorporates two key technologies, SCOUT and ERL, significantly improving learning efficiency and task execution stability.