StepFun has launched a new family of on-device models called Step Edge, designed to run directly on phones and cars. The lineup includes a base model, Audio, GUI, and Gen, handling text, images, and speech locally. It can understand screens, recognize speech, control interfaces, and even generate images. The company claims tool call latency can be as low as 0.1 seconds.
Simple, frequent, or offline tasks stay on the device, while complex reasoning is sent to the cloud. That means sensitive data never has to leave your phone, and you're less dependent on network speed and cloud computing power.
StepFun also debuted its own Step Inference NPU inference engine, optimized for terminal chips to cut latency further. The company says Step Edge leads in 29 core benchmark metrics.