Huawei is squeezing a 30B-parameter on-device MoE model into its new tri-fold Mate XT 2. It's a mixture-of-experts design, so only 2B of those 30B parameters are active at a time, and Huawei says that makes it the industry's first on-device, omni-modal MoE large model. The model runs on the freshly announced Kirin 9050 Pro, with the Da Vinci-architecture NPU providing the on-device compute.
The 30B/2B on-device model was actually teased back in June. Huawei said then that it was specifically optimized for its Kirin silicon, using techniques like expert prediction to lower memory overhead and improve inference throughput, on track for a fall arrival in Kirin-powered chips and phones.