Skip to main content

Step Astronomy's New Edge AI Models Run Fast on Your Phone

Step Astronomy has unveiled its latest Step Edge series of terminal models, bringing powerful multimodal AI capabilities directly to smartphones, in-vehicle systems, and other edge devices. The lineup includes four variants: Basic, Audio, GUI, and Gen, each tailored for specific tasks like screen understanding, speech recognition, interface control, and even image generation.

What sets these models apart is their ability to process text, images, and audio entirely on the device. According to the company, the call latency is as low as 0.1 seconds, making interactions feel nearly instantaneous. For simple, frequent tasks or in weak network conditions, everything happens locally. Complex reasoning, on the other hand, can be offloaded to the cloud when needed. This hybrid approach significantly cuts down on network and cloud dependency, and because sensitive data never leaves the device, privacy and security get a major boost.

Image

To squeeze even more performance out of edge hardware, Step Astronomy also introduced its self-developed Step Inference NPU inference engine. This engine is designed to dramatically reduce inference latency on terminal chips, ensuring that even demanding AI tasks run smoothly.

The company claims that the Step Edge series leads in 29 core evaluation metrics, showcasing state-of-the-art multimodal performance on the terminal side. This move signals a further acceleration of AI terminalization, promising more real-time and private user experiences for smart devices like phones and cars.

Key Points

  • Four model variants: Basic, Audio, GUI, and Gen for different local tasks.
  • Ultra-low latency: 0.1-second response time for local processing.
  • Hybrid architecture: Simple tasks run locally; complex ones use the cloud.
  • Privacy first: Sensitive data stays on the device.
  • Custom NPU engine: Step Inference NPU boosts terminal chip performance.
  • Top benchmarks: Leads in 29 core evaluation metrics.