February 2026 monthly summary for jd-opensource/xllm. Delivered Torch NPU MoE kernel optimizations, introducing grouped MatMul, gating softmax, and routing initialization. This feature enhances performance and scalability of Mixture of Experts workloads on NPU hardware, enabling larger expert pools and faster inference. Work committed under fa67f078d3cb4ec8f39dfd14fe0435e31cf19e63 and merged as part of PR #924, with co-authors shenxiaolong and ext.wangxiaochi1. This lays groundwork for broader NPU acceleration, demonstrates cross-team collaboration, and strengthens the repository’s readiness for future MoE enhancements. Overall, the month focused on feature delivery with tangible performance/throughput benefits. No major bugs reported in this period; primary value delivered comes from performance optimization and deeper NPU integration.