Step Function发布Step Edge, reducing End-to-End Agent tool invocation time to as low as 0.1 seconds

According to Perceive Beating monitoring, Step Edge, a flagship release from LeapStar, has introduced its on-device model family, which includes base models for Audio, GUI, and Gen, targeting mobile phones, automobiles, and other terminals. It is capable of locally processing text, images, and speech to achieve screen understanding, speech recognition, interface operation, and image generation. The on-device tool has a minimal call latency of 0.1 seconds.

Simple, high-frequency, or low-network tasks can be completed locally, while complex inference can be offloaded to the cloud. Sensitive data does not need to be uploaded, reducing reliance on network and cloud computing power.

LeapStar has also launched its self-developed Step Inference NPU Inference Engine, aimed at reducing latency for terminal chips. The company claims that the Step Edge series has ranked first in 29 core evaluation metrics.