Moore Threads Technology Co.,Ltd. announced today that it has successfully achieved comprehensive adaptation of Alibaba's latest large language model, Qwen3.5, on its flagship AI training and inference GPU, the MTT S5000. This milestone demonstrates the maturity and robustness of the company's MUSA ecosystem, enabling developers to efficiently deploy and optimize models using the MUSA C programming language and the Triton-MUSA toolchain.
During the adaptation process for Qwen3.5, Moore Threads validated two core capabilities of the MUSA ecosystem. Native MUSA C support allows developers to perform kernel development directly using MUSA C, significantly lowering the barrier to migration from the CUDA ecosystem. Deep compatibility with Triton-MUSA enables developers to write high-performance operators using the familiar Triton syntax and run them seamlessly on Moore Threads' full-feature GPUs via the Triton-MUSA backend.
For the hybrid attention mechanism employed by the multimodal Qwen3.5 model, Moore Threads implemented native optimizations. Leveraging the muDNN computing library and the MATE open-source operator library, the company provides efficient support for long-sequence processing within the hybrid attention mechanism, achieving high-performance inference for the model on the MTT S5000. This achievement not only reaffirms the broad adaptability and efficient support capabilities of domestic full-feature GPU computing platforms for cutting-edge large models but also highlights the significant benefits of hardware-software co-optimization.
From GLM-5 to MiniMax M2.5, and now to Qwen3.5, Moore Threads has consistently maintained rapid tracking and adaptation of leading domestic large models. This agile response mechanism stems not only from the MUSA architecture's seamless compatibility with mainstream AI ecosystems and its continuously optimized toolchain support but also signifies that the domestic computing infrastructure now possesses end-to-end support capabilities, from model adaptation to efficient deployment. In the future, Moore Threads will continue to deepen its MUSA technology ecosystem, leveraging a more robust and user-friendly domestic computing foundation to help more cutting-edge large models achieve rapid deployment and accelerate the prosperity of the domestic computing ecosystem.
Comments