Moore Threads MTT S5000 Now Supports MiniMax M3 Model

Moore Threads has announced that its flagship AI training and inference card, the MTT S5000, now supports MiniMax's new multimodal model M3 from day one. The adaptation covers hardware, software, and open-source frameworks.

For M3's long context feature, the MTT S5000 uses native FP8 acceleration to deliver 1000 TFLOPS of AI power, with 80GB of memory and 1.6TB/s bandwidth for million-token sequences. New operators can be quickly migrated via MUSA C++ and Triton-MUSA layers.

For M3's coding and agentic tasks, Moore Threads optimized inference throughput and latency using custom operators, and launched support for vLLM and SGLang frameworks. The MTT S5000 supports FP8 to FP64 precision, covering the full development-to-deployment pipeline.