AMD Adds MiniMax-M3.1 Support for Instinct MI350 Series GPUs

AMD announced Thursday, Oct. 22, that MiniMax-M3.1 can run on Instinct MI350 Series GPUs through ROCm versions of vLLM and ATOM, bringing a coding and agent model with a context window of one million tokens to AMD’s inference stack.

MiniMax-M3.1 includes reasoning and tool-use features. AMD’s setup supports the original NVFP4 checkpoint after conversion during loading, as well as an MXFP4 export made with AMD Quark. The ATOM configuration AMD tested uses the Quark export. AITER kernels handle expert routing, attention and reduction operations. AMD said the two weight paths performed within evaluation error of each other on GSM8K, AIME and MMMU-Pro.

ATOM applies per-channel FP8 attention and combines shared-expert work during startup. It also uses CUDA graphs to capture DSpark’s draft pass; AMD’s tests found that roughly half of DSpark’s seven-token draft proposals were accepted on real text. The MI350 implementation uses four-way tensor parallelism per replica, allowing two replicas on an eight-GPU node.

In AMD’s tests, vLLM mono decode reduced per-token decode latency by a factor of 3.2 at one request, 2.1 at four and 1.5 at 16, compared with the same server with mono decode disabled. On coding-agent traces, AMD said mono decode produced 2.2x to 2.4x the interactivity across one to eight concurrent users. Pairing it with DSpark speculative decoding increased interactivity by about one-third for one to four users and 13% at eight users in the reported tests.

The current configuration supports mono decode with DSpark for up to four concurrent requests. Wider decode steps use the standard path. AMD listed wider mono-decode steps and support for MiniMax-M3.1 mono decode in ATOM as future work, without a timeline.

In a separate announcement on Oct. 12, AMD introduced its Helion rack-scale design for agentic AI workloads; its enterprise configuration supports up to 16 Instinct MI350P GPUs.