Moore Threads has completed day-zero adaptation for Moonshot AI's Kimi K3.1 Pro on its MTT S5000 AI training and inference card and MUSA software stack. According to a company announcement reported by IT Home on November 6, the work connects model loading, service startup, and basic text and image inference.
The company characterized the achievement as a functional deployment rather than a published performance comparison. Moore Threads stated it migrated the model's key inference paths and enabled stable execution on the MTT S5000. However, the report did not provide specific data regarding throughput, latency, power consumption, or the number of cards used.
The day-zero description is the company's own characterization, following Moonshot's release of the full model weights on November 4. At the framework level, the team utilized SGLang-MUSA to build on the upstream model implementation. The migration process also involved the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler.
The MUSA stack covers framework execution, operator compilation, device runtime, and communication across multiple cards. This allows developers to reuse existing interfaces while adapting new computation paths. Much of the model-specific work focused on memory handling and attention. For DSA sparse attention, the stack adapts index selection, key-value cache layout, and both prefill and decoding.
Moore Threads stated that MATE sparse-attention operators work alongside Kimi Delta Attention state management, cross-layer residual connections, and visual attention to form the execution chain.
The team also addressed the deployment of the model's MXFP4 expert weights. Instead of generating a complete offline FP8 copy before serving, the team converts weights online and reuses an FP8 workspace across layers for expert computation. Moore Threads presented this as a method to bring the very large model into service without first preparing a separate full set of converted weights.
