Physix Frontier badgePhysix Frontier
PublishChinese | EN

Physix Frontier · News Briefing Card (IT Home · Oct 13, 2026)

Huawei Unveils Kirin 9050 Series, First 30B MoE LLM On-Device

KEY FACTS

  • Huawei held its 2026 Kirin media briefing in Shanghai on October 13, where Xu Songyu delivered a keynote.
  • The Kirin 9050 series is the first to apply the Tao Law and logic folding approach, yielding gains in performance and energy efficiency.
  • Huawei aims to push P-core frequency past 5GHz by 2031.
  • The Kirin 9050 Pro packs a next-generation Da Vinci architecture NPU and Lingxi micro-core, boosting AI performance by 140%.
  • The Kirin 9050 series achieves the industry's first on-device 30B MoE large model, using less than 2.5G of memory.

KEY DATA

140%AI performance uplift
66%AI power reduction
3BMoE total parameters
2BMoE activated parameters

PHYSIX OBSERVATION

A 30B MoE running on-device with under 2.5G of memory shows that on-device large models are moving from demos to practical use, and running large models locally on phones no longer depends on the cloud. Huawei stretching its chip roadmap to 2031 is using a long-cycle rhythm to hedge against process constraints. What's worth watching is that the on-device AI compute race will kick off early.

Source: IT Home report