AMD Brings MiniMax-M3.1 Inference Support to Instinct MI350 GPUs
AMD has added support for the MiniMax-M3.1 artificial intelligence model to its Instinct MI350 Series accelerators through ROCm-enabled versions of vLLM and ATOM. The update gives developers access to a coding and agent-focused model with a context window of up to one million tokens.
MiniMax-M3.1 combines reasoning capabilities with tool use, making it suitable for coding assistants and other agentic workloads. AMD’s implementation supports the model’s original NVFP4 checkpoint, which is converted during loading, as well as an MXFP4 version produced with AMD Quark. The configuration used in AMD’s testing relied on the Quark-generated model.
Quantization and kernel support
AMD’s implementation uses AITER kernels for key operations, including expert routing, attention and reductions. According to the company’s evaluations, both supported weight formats produced results within the expected evaluation margin on GSM8K, AIME and MMMU-Pro benchmarks.
ATOM also applies per-channel FP8 attention and combines shared-expert processing during startup. Its implementation uses CUDA graphs to capture the draft phase of DSpark, a speculative decoding system designed to predict several tokens before the main model confirms them.
In testing with real-world text, approximately half of DSpark’s seven-token draft sequences were accepted. The MI350 setup uses four-way tensor parallelism for each replica, allowing two replicas to run across an eight-GPU server.
Lower latency and improved interactivity
AMD reported notable improvements when using mono decode with vLLM. In single-request testing, per-token decode latency was reduced by a factor of 3.2 compared with the same server operating without mono decode. At four requests, the improvement was 2.1 times, while a 16-request workload delivered a 1.5-times reduction.
The company also measured performance using coding-agent workloads. Mono decode increased interactivity by between 2.2 and 2.4 times across tests involving one to eight simultaneous users.
Adding DSpark speculative decoding provided another performance boost. Interactivity rose by approximately one-third in tests with one to four users, while the improvement reached 13% with eight concurrent users.
Current limitations and future plans
The present configuration supports mono decode combined with DSpark for as many as four concurrent requests. Larger decode batches continue to use the standard processing path.
AMD identified broader mono-decode steps and MiniMax-M3.1 mono-decode support in ATOM as future development goals, but did not provide a schedule for their arrival.
The announcement follows AMD’s introduction of its Helion rack-scale architecture for agentic AI workloads. The enterprise version of that platform is designed to support systems equipped with up to 16 Instinct MI350P GPUs, positioning the accelerator family for increasingly demanding inference deployments.
