Huawei has spelled out the artificial-intelligence hardware inside its next flagship phone silicon, saying the neural engine powering the Kirin 9050 series can run a large language model with roughly 30 billion parameters without sending data to the cloud. The company’s Kirin unit published the technical details on November 11, describing the chip as its first built with a "logic folding" design that packs more compute into the tight power budget of a smartphone.
At the center of the announcement is a new generation of Huawei’s in-house Da Vinci architecture, the neural processing unit the company uses for on-device AI. Huawei said the redesigned engine delivers 140 percent higher performance while cutting power consumption by 66 percent compared with the previous generation. It includes a dedicated transformer engine, an ultra-low-power sensing block and support for the CANN heterogeneous computing stack that Huawei uses across its Ascend data-center chips and its consumer devices.
The headline claim is on-device support for a 30-billion-parameter mixture-of-experts model, which Huawei described as an industry first at this scale inside a handset. Fitting a model that large onto a phone relies on multi-level coordination between memory and compute plus aggressive low-bit quantization, techniques that shrink the model’s memory footprint and reduce the energy each answer costs. Huawei said the setup works offline, so common tasks no longer depend on a stable connection or a round trip to a remote server.
The company also showed what the hardware means in daily use. The sensing engine can read screen content aloud, power an upgraded version of the Xmage computational photography system, turn pages by tracking eye movement and keep turn-by-turn navigation alive on a screen-off map. Each feature draws on a different combination of vision, language and always-on sensing rather than treating AI as a single chatbot feature.
The Kirin 9050 line was first previewed at a media briefing in October, where Huawei positioned it as the opening of a decade-long roadmap for the logic-folding approach. A company slide shown at the event set a target of pushing the chip’s performance cores past 5 gigahertz by 2031. Chinese phone makers have spent two years racing to run ever-larger models locally, arguing that on-device processing is faster, cheaper at scale and more private. Huawei’s new claim sharpens that race just as its domestic rivals prepare their own flagship launches for the year-end shopping season.

Dgcity