Huawei Kirin 9050 series is equipped with a new-generation Da Vinci architecture NPU, supporting on-device operation of 30B large models
Huawei Kirin officially detailed the on-device AI capabilities of the Kirin 9050 series. This series is equipped with a new-generation Da Vinci architecture NPU and can support a 30B MoE large model running offline on the phone, completing complex task planning, intelligent photo editing, and multimodal perception. The NPU performance of the Kirin 9050 Pro is 140% higher than the previous generation, with power consumption reduced by 66%, and it integrates the Transformer engine, ultra-low-power perception engine, and CANN heterogeneous computing architecture. The series also supports a 6B image editing model, a 7B Dense text model, and a small-sized multimodal perception model. Among them, the 30B MoE model has a total parameter count of 30 billion, an activated parameter count of about 2 billion, and additional memory usage of less than 2.5GB.
Why read this
SelectedAI score 85
The original text provides specific figures for NPU performance improvement and power consumption reduction, as well as the model specifications that can run on-device, making it easy to judge the capability boundaries of local AI on phones.
Topics
Body · AI translation
Huawei Kirin 9050 Series Officially Announced with New-Generation Da Vinci Architecture NPU: Supports 30B Large Model Running On-Device Phoenix Tech Download the App for Exclusive Early Access November 11, 2026 12:02:52 From Beijing
Phoenix Tech News (Editor/Song Shumei) On November 11, Huawei Kirin officially published an article detailing the on-device AI capabilities of the Kirin 9050 series. The series is equipped with a new-generation Da Vinci architecture NPU and can support a 30B MoE large model running on mobile phones, completing complex task planning, intelligent photo editing, and multimodal perception without connecting to the network or waiting for cloud processing.
According to data released by Huawei, the NPU performance of the Kirin 9050 Pro is 140% higher than the previous generation, while power consumption is reduced by 66%. Its NPU integrates a Transformer engine, an ultra-low-power perception engine, and the CANN heterogeneous computing architecture, providing computing power support for on-device models of different scales.
The Kirin 9050 series adopts a collaborative architecture of an on-device "brain" and "cerebellum." The "brain" is responsible for understanding user intent, performing complex reasoning, and multi-step task planning; the "cerebellum" handles always-on low-power perception and rapid response tasks, enabling AI to move further from recognizing instructions to understanding intent.
In terms of model deployment, the series supports a 30B MoE task planning model, a 6B image editing model, a 7B Dense text model, and small-size multimodal perception models. Huawei previously disclosed that the 30B MoE model has a total parameter count of 30 billion, an activated parameter count of about 2 billion, and additional memory usage of less than 2.5GB; the 7B text model improves first-token response performance by 180% and reduces model inference power consumption by 29%.
Based on the above capabilities, terminals equipped with the Kirin 9050 series can complete multi-step task processing, offline intelligent photo editing, and text generation locally, and can also continuously perceive screen, image, human eye, and location information, providing on-device computing power support for features such as AI screen reading, smart page turning, and screen-off maps.