Moore Threads completes full adaptation of Qwen3.5 model
I'm LongbridgeAI, I can summarize articles.Chinese GPU maker Moore Threads has completed the full adaptation of Alibaba's Qwen3.5 model on its MTT S5000 graphics processor. This adaptation includes compatibility across the entire pipeline—training, inference, and quantized deployment—supporting various precision formats like FP16, BF16, and INT4. The MUSA ecosystem allows developers to utilize the MUSA C programming language and Triton-MUSA toolchain for efficient model optimization and deployment. Additionally, enhancements in long-sequence processing via the muDNN computing library have improved inference performance.
Chinese GPU maker Moore Threads said it has completed full adaptation of Qwen3.5, the latest open-source large language model from Alibaba, on its flagship MTT S5000 graphics processor. The company said the compatibility spans the full pipeline, including training, inference and quantized deployment, with support for multiple precision formats such as FP16, BF16 and INT4. Built on its MUSA ecosystem, Moore Threads said developers can use the native MUSA C programming language and the Triton-MUSA toolchain to optimize and deploy models more efficiently. To support Qwen3.5’s hybrid mechanism, the company said it enhanced long-sequence processing through its muDNN computing library, leading to improved inference performance. [TechNode Reporting]