Back to news
launchMoore ThreadsMoonshot AI2026-11-06

Moore Threads announced on November 6 the Day-zero adaptation of the 3 trillion parameter Kimi-K3.1 Pro

Moore Threads used its MTT S5000 AI training and inference card with the MUSA software stack to enable model loading, service startup and basic text and image inference for Kimi-K3.1 Pro, including sparse attention and MXFP4 expert weights.

On November 6, Moore Threads announced that, based on its MTT S5000 AI training and inference all-in-one computing card and the MUSA software stack, it completed Day-zero adaptation of the Kimi-K3.1 Pro model, enabling model loading, service startup and basic text and image inference paths.

Kimi K3.1 Pro is the first model in the K3.1 series and the most capable 3 trillion parameter model so far, supporting a 1 million token context. After large-scale post-training and reinforcement learning, the model improved across code, three-dimensional creation, financial research, knowledge question answering and long-horizon agent tasks, reaching or approaching top closed-model levels on code and agent benchmarks, with 65.7 percent on Terminal-Bench 4.0 and 61.2 percent on the Kimi Code Bench 3.0 long-horizon task, above the 55.95 percent of K3.

Around the model architecture and computational characteristics of Kimi-K3.1 Pro, Moore Threads completed key inference path migration and adaptation through the MUSA software stack. The team took over the upstream model implementation on SGLang-MUSA and completed migration and integration of critical new-architecture compute paths through the MATE operator library, Triton-MUSA, TileLang-MUSA and the MTCC compiler. For the model's sparse attention DSA, the MUSA software stack adapted index selection, KV cache layout and the Prefill and Decode compute paths. For MXFP4 expert weights, the team executes expert computation through online conversion and cross-layer reuse of an FP8 workspace, without needing to pre-generate complete offline FP8 weights.

摩尔线程Kimi K3.1 ProDay-0 适配MUSA国产算力