December 2025 monthly summary for vllm-ascend: Delivered a new custom operator aclnnMoeInitRoutingCustom for Mixture-of-Experts (MoE) models, enabling an initialization-routing path that can be replaced by aclnnMoeInitRoutingV3 once CANN 8.5 is available. The change was implemented in vllm-project/vllm-ascend via a single PR (commit d5f72835e65e6b6b8cf8e173a9048c926890881b) and CI tests passed, including new and existing tests. This work lays groundwork for scalable MoE routing on Ascend hardware, with a modular operator design that supports future performance optimizations and hardware-specific acceleration. There are no user-facing API changes in this release; the focus is on internal routing initialization and upgrade readiness. The effort demonstrates strong collaboration, code quality, and a maintainable path toward MoE routing improvements in the Ascend ecosystem.