Huawei's Kirin 9050 series brings a new Da Vinci NPU capable of running 30B MoE models on device
Huawei detailed the new Da Vinci NPU in the Kirin 9050 series on November 11, claiming 140 percent higher performance and 66 percent lower power draw, with on-device support for a 30B MoE model without network or cloud round trips.
Huawei's Kirin official account published details on November 11 about the new Da Vinci NPU in the Kirin 9050 series, the first chip built on its logic folding technology. According to the official description, the NPU supports on-device execution of 30B-scale models without needing a network connection or waiting on the cloud.
On performance, the new Da Vinci NPU delivers 140 percent higher performance and 66 percent lower power consumption, while supporting a Transformer engine, an ultra-low-power sensing engine and the CANN heterogeneous computing architecture, with multi-level compute-in-memory co-optimization and low-bit quantization. The company describes it as the first solution to bring a 30B MoE model on device. Concrete functions include screen reading for screen sensing, image capture 2.0 for imaging sensing, eye-gaze page turning for eye tracking, and screen-off maps 2.0 for location sensing.
Earlier, at the 2026 Kirin media communications conference held in October, Xu Songyu, head of wireless terminal chip product management in Huawei's semiconductor business, said the Kirin 9050 series as the first logic folding chip would open a new ten-year journey, with the official target of pushing performance-core frequency past 5GHz by 2031. Being able to carry a 30B mixture-of-experts model on device means voice assistants and image processing on phones no longer depend entirely on cloud inference, though sustained power performance in real workloads still needs validation on actual devices.