Huawei Unveils Kirin 9050 Series Enabling Industry's First On-Device 30B MoE Model
At the 2026 Kirin Media Communication Conference, Huawei introduced the Kirin 9050 series featuring the new 'Tao Law' and logic folding architecture. The platform achieves a major milestone by running the industry's first on-device 30B Mixture of Experts (MoE) large language model. This announcement marks a significant leap in mobile AI capabilities, allowing smartphones to run heavy, highly capable language models locally with high efficiency. It highlights Huawei's ongoing semiconductor innovation despite external constraints, setting a new performance benchmark for mobile computing. The Kirin 9050 Pro is equipped with a new generation DaVinci architecture NPU and Lingxi micro-core, delivering a 140% increase in AI performance while reducing power consumption by 66%. The on-device 30B MoE model features 30B total parameters, 2B activated parameters, and consumes less than 2.5GB of memory.
## BACKGROUND
Mixture of Experts (MoE) is a machine learning architecture that activates only a subset of parameters for any given task, balancing high model capacity with lower computational costs. Logic folding is an advanced circuit design approach that stacks 2D circuit logic paths into 3D configurations to enhance physical signal transmission, efficiency, and performance.