Kirin 9050 Series Brings 30B MoE Models On-Device
Huawei announced that the Kirin 9050 series uses a new Da Vinci-architecture NPU that can run 30B MoE models on-device without a network connection or cloud processing. Huawei claims the NPU delivers 140% higher performance and 66% lower power consumption, with Transformer engine, low-power sensing, CANN heterogeneous computing, multi-level compute-memory optimization, and low-bit quantization support. The vendor also listed AI screen reading, Xmage 2.0 photography, eye-controlled page turning, and screen-off maps as possible features.
The announcement positions 30B-class MoE inference on a smartphone chip, potentially expanding offline AI capabilities, although the performance and power figures remain Huawei claims relayed by IT之家.