Huawei has detailed how the Kirin 9050 series chip runs a 30-billion-parameter AI model entirely on the device without a cloud connection. The announcement came through the brand’s official channel and was reported by outlets such as IT Home and TechNode on November 11, 2026.
What happened?
The Kirin 9050 is Huawei’s first chip with “logic folding” (τ) technology. Its new-generation NPU based on the Da Vinci architecture delivers 140% higher AI performance and reduces power consumption by 66% compared with the previous generation. It supports the Pangu 30B-A2B model, a Mixture-of-Experts (MoE) design that activates roughly 2 billion parameters at a time, using low-bit quantization and multi-level memory-compute coordination.
How it works in practice
By not activating all 30 billion parameters at once, the model fits within a smartphone’s power, heat and memory limits. The NPU includes a Transformer engine, ultra-low-power perception engine and CANN heterogeneous computing architecture. Mentioned features include AI screen reading, Xmage 2.0 intelligent capture, eye-tracked page turning and always-on maps.
Why it matters
According to Huawei, this is the first 30B MoE model running locally on a phone chip. It reduces reliance on cloud servers, improves privacy and latency, and marks a step forward for on-device AI. The company had already previewed some of the numbers in October during a technical briefing.
Context
The Kirin 9050 is part of Huawei’s LogicFolding chip line and powers devices such as the Mate series. The company’s long-term goal includes performance cores above 5 GHz by 2031. The release reinforces China’s strategy for semiconductor and embedded-AI sovereignty.
Image credit: Huawei | Source: IT Home / TechNode
By Grok
