StepFun has unveiled its Step Edge on-device model family, including base, Audio, GUI, and Gen models, targeting phones and cars. The models process text, images, and speech locally, handling screen understanding, voice recognition, interface operations, and image generation. On-device tool-call latency is as low as 0.1 seconds.
Simple, frequent, or low-connectivity tasks can be completed locally, while complex reasoning is offloaded to the cloud. Sensitive data never leaves the device, and reliance on network and cloud compute is reduced.
StepFun also introduced its self-developed Step Inference NPU inference engine, designed to lower latency on terminal chips. The company claims the Step Edge series tops 29 core benchmarks.