Physix Frontier · News Briefing Card (IT Home · Nov 5, 2026)

Huawei Ascend completes day-one inference support for Kimi K3.1 Pro

KEY FACTS

  • Moonshot AI released the full model weights of Kimi K3.1 Pro on November 4, 2026.
  • Ascend completed inference adaptation on launch day through vLLM Ascend and SGLang.
  • The adaptation covers Atlas 800 A3, Atlas 900 A3 SuperPod, and Ascend 950 super nodes.
  • Kimi K3.1 Pro has 2.8 trillion parameters and natively supports vision understanding and a one-million-token context.
  • Ascend completed operator adaptation for the newly added sparse attention path and multimodal inference.

KEY DATA

2.8 trillionModel parameter scale
one million tokensContext window

PHYSIX OBSERVATION

Day-one synchronized adaptation between domestic compute and leading large models is becoming the norm, and Ascend is trading 0 Day support for developer mindshare. For users, the barrier to deploying ultra-large MoE models is falling; but operator optimization for sparse attention and long context remains the decisive factor for performance, and actual deployment efficiency is yet to be verified.

Source: IT Home report