Huawei is making its own neural processor the selling point of Kirin 9050 phones, handling task planning, photo editing and screen reading on the device without a cloud connection.
Huawei says its Kirin 9050 phone chips can run a 30-billion-parameter AI model entirely on the handset, with no network connection and no wait for a cloud server. IT Home reported the details on November 11 from a post on Huawei’s official Kirin WeChat account, which credited a new-generation Da Vinci neural processing unit (NPU), the part of the chip built for AI work, along with low-bit quantization and coordinated use of memory and compute. TechNode described the model as a mixture-of-experts design that activates about 2 billion parameters at a time.
According to Huawei, the new NPU is 140% faster than the previous generation and uses 66% less power. It combines a Transformer engine, an ultra-low-power perception engine and Huawei’s CANN heterogeneous computing architecture. The features Huawei tied to it are AI screen reading, Xmage smart capture 2.0, AI eye-tracking page turning and a screen-off map 2.0.
The model size and the 2-billion figure were first disclosed in October. What is new is Huawei’s account of how the model fits in a phone and which tasks it is meant to handle. All of the figures come from Huawei, and no independent measurement of speed, memory use or battery drain has been published.
Huawei splits on-device AI into a large planner and smaller helpers
A Phoenix report republished by 10jqka set out the lineup. The 30-billion-parameter mixture-of-experts model handles task planning, with about 2 billion parameters active and less than 2.5GB of added memory. It sits beside a 6-billion-parameter image-editing model, a 7-billion-parameter dense text model and a small multimodal perception model. Huawei says the 7B model’s first response arrives 180% faster and its inference uses 29% less power.
Huawei describes the arrangement as a "brain" and a "cerebellum". The brain handles intent, complex reasoning and multi-step planning; the cerebellum stays on at low power to sense the screen, images, the user’s eyes and location, and reacts quickly. A mixture-of-experts model routes each request to a small set of specialized sub-networks, which is how a 30-billion-parameter model can run with only a fraction of its weights working at once.
Huawei says the 30-billion-parameter model on its Kirin 9050 chips activates about 2 billion parameters at a time and needs less than 2.5GB of extra memory. Low-bit quantization, which stores each weight with fewer bits, is one of the techniques Huawei credits for that figure, although it has not said how many bits it uses.
The claims build on an October briefing and the Mate 90 launch
Huawei first presented the model at its Kirin media briefing on October 13 at its Lianqiu Lake site in Shanghai. IT Home’s briefing report said Xu Songyu, head of product management for wireless terminal chips in Huawei’s semiconductor business, presented the Kirin 9050 series as the first chip built on Huawei’s "logic folding" approach, alongside the 30B model. IT Home’s November report quoted Xu saying the series "will open a new journey for the next 10 years". A slide shown there targeted CPU performance-core frequencies above 5GHz by 2031.
Two weeks earlier, at the Mate 90 launch on October 1, Huawei consumer chief Richard Yu presented the Kirin 9050 Pro with transistor density of 238 million per square millimeter, up 28%, and a demonstration of offline AI photo editing running on the NPU, CNMO reported. The 9050 Pro debuted in September in the Mate XT 2 trifold. Huawei’s Mate 90 Pro Max specifications list both the Kirin 9050 Pro and the Kirin 9050, with 12GB or 16GB of memory, and the 9050 Pro alone for the collector’s edition.
Independent tests will decide whether the model is useful
Huawei has published capability claims, not usage data. It has not released the model’s response speed in tokens per second, the bit width of its quantization or the battery cost of keeping the perception engine running. Those figures will determine whether a 30-billion-parameter planner is a practical assistant or a showcase.
The test to watch is third-party measurement on shipping Mate 90 Pro Max units: whether offline task planning and eye-tracking page turning run at usable speed on the 12GB configuration as well as the 16GB one.