Huawei detailed how the Kirin 9050 Series chip runs a 30-billion-parameter MoE language model locally without needing an internet connection.
What happened?
On November 11, 2026, Huawei published details about the new generation of the Da Vinci NPU architecture on the Kirin 9050, the company's first chip with “logical folding” (τ) technology. The NPU delivers 140% more performance and 66% lower power consumption compared to the previous generation. It is the first in the industry to allow a 30B-parameter Mixture-of-Experts model to run entirely on-device.
The model activates about 2 billion parameters at a time. Huawei uses low-bit quantization and coordination between memory and compute to keep everything local. Practical features include AI screen reading, Xmage 2.0 smart capture, eye-tracked page turning, and offline maps 2.0.
Why does this matter?
Running a 30B model locally on a smartphone represents a significant leap in edge AI. It reduces cloud dependency, improves privacy and latency. Huawei positions the Kirin 9050 as the start of a decade of evolution based on the “τ law,” which prioritizes signal latency reduction rather than pure geometric transistor shrinkage. Transistor density rose 53.5%.
The chip integrates Linxi CPU, Maleoon GPU, and the new NPU, plus Balong modem. Long-term goals include P-cores above 5 GHz by 2031.
What changes in practice?
Smartphones with Kirin 9050 will be able to run AI assistants, photo organization, and multimodal features offline. Users get faster responses and greater privacy. For the industry, it shows that high-performance Chinese chips for local AI are advancing rapidly, even under manufacturing constraints.
There is still no device launch date, but the technical details have already been officially released.
By Grok
