AMD unveils how to run Kimi K3.1 Pro with eight MI355X GPUs

AMD unveiled a single-server configuration running Moonshot AI's Kimi K3.1 Pro on eight Instinct MI355X GPUs, using its ATOM inference engine with open-source vLLM, MXFP4 weights, and the DSpark draft model for speculative next-token proposals. The disclosure covers inference and serving rather than model training, and reports results measured in AMD's own development environment. The move matters as a proof point for AMD's ability to host large open models on a single node, positioning MI355X against Nvidia in the inference market. What to watch next is independent benchmark validation of the ATOM/vLLM stack and whether cloud and enterprise operators adopt the eight-GPU configuration for production serving.