Moore Threads Completes Day-0 Adaptation for 3-Trillion-Parameter Kimi-K3.1 Pro
Moore Threads announced Day-0 adaptation of Moonshot AI's 3-trillion-parameter Kimi-K3.1 Pro multimodal model on its MTT S5000 AI accelerator and MUSA software stack, successfully establishing text and image inference pipelines. This milestone demonstrates the growing capability of domestic AI hardware and software stacks to support massive, cutting-edge open-weight models simultaneously with their release. It accelerates the integration of high-performance localized AI infrastructure with advanced large language models. The adaptation utilizes SGLang-MUSA, the MATE sparse attention operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler to handle complex model architectures and MXFP4 expert weights efficiently. Kimi-K3.1 Pro features a 1-million-token context window and exceptional performance in coding and agent tasks.
## BACKGROUND
Moore Threads is a Chinese GPU manufacturer that develops full-function graphics and AI computing accelerators powered by its proprietary MUSA software stack. The MTT S5000 is a high-performance universal GPU designed for large-scale AI training, fine-tuning, and inference workloads.