Moore Threads Completes Full Adaptation of Alibaba's Qwen3.5 Model on MTT S5000 GPU
Chinese graphics processing unit manufacturer Moore Threads has announced the successful full adaptation of Qwen3.5, Alibaba Group's latest open-source large language model, on its flagship MTT S5000 graphics processor. This strategic integration ensures comprehensive compatibility across the entire development pipeline, encompassing model training, inference, and quantized deployment. The adaptation supports multiple precision formats, including FP16, BF16, and INT4, providing developers with flexible options for optimizing performance and resource usage. Leveraging its proprietary MUSA ecosystem, Moore Threads enables developers to utilize the native MUSA C programming language and the Triton-MUSA toolchain for efficient model optimization and deployment. To specifically address the hybrid mechanisms inherent in Qwen3.5, the company enhanced long-sequence processing capabilities through its muDNN computing library, resulting in significantly improved inference performance. This development highlights the growing synergy between domestic hardware manufacturers and leading AI software providers in China, aiming to strengthen the local artificial intelligence infrastructure. By ensuring seamless integration between advanced large language models and domestic GPU hardware, Moore Threads aims to facilitate more robust and efficient AI applications for enterprise and developer communities utilizing their hardware solutions.
Wire timeline
Moore Threads Completes Full Adaptation of Alibaba's Qwen3.5 Model on MTT S5000 GPU
Chinese graphics processing unit manufacturer Moore Threads has announced the successful full adaptation of Qwen3.5, Alibaba Group's latest open-source large language model, on its flagship MTT S5000 graphics processor. This strategic integration ensures comprehensive compatibility across the entire development pipeline, encompassing model training, inference, and quantized deployment. The adaptation supports multiple precision formats, including FP16, BF16, and INT4, providing developers with flexible options for optimizing performance and resource usage. Leveraging its proprietary MUSA ecosystem, Moore Threads enables developers to utilize the native MUSA C programming language and the Triton-MUSA toolchain for efficient model optimization and deployment. To specifically address the hybrid mechanisms inherent in Qwen3.5, the company enhanced long-sequence processing capabilities through its muDNN computing library, resulting in significantly improved inference performance. This development highlights the growing synergy between domestic hardware manufacturers and leading AI software providers in China, aiming to strengthen the local artificial intelligence infrastructure. By ensuring seamless integration between advanced large language models and domestic GPU hardware, Moore Threads aims to facilitate more robust and efficient AI applications for enterprise and developer communities utilizing their hardware solutions.
TechNode