News
AMD says MiniMax-M3.1 now runs on Instinct MI350 GPUs with vLLM and ATOM support, MXFP4 conversion and latency gains in select decode paths.
AMD said on Oct. 22, 2026 that MiniMax-M3.1 now has day-0 support on Instinct MI350 Series GPUs across its ROCm inference stack, including the ROCm builds of vLLM and ATOM. The company said the natively multimodal MoE model can run with either the original NVFP4 checkpoint converted at load time or an MXFP4 export produced with AMD Quark, with separate kernel and serving paths for latency-bound and higher-concurrency deployments.
Evidence and context
In AMD’s technical article published Oct. 22, 2026, the company described MiniMax-M3.1 as a model for coding and agent workloads with a 1M-token context, built-in reasoning, native tool calling and DSpark speculative decoding. AMD said both vLLM and ATOM serve the model on MI350 GPUs, while AITER kernels handle the routed experts, attention and reduction paths. The article also said the MXFP4 checkpoint produced with AMD Quark, and an equivalent conversion of the NVFP4 release checkpoint at load time, matched within evaluation error on GSM8K, AIME and MMMU-Pro.
AMD reported that its vLLM mono-decode path cut per-token decode latency by 3.2x at one request, 2.1x at four requests and 1.5x at sixteen requests versus the same server with mono decode disabled. On real coding-agent traces, AMD said mono decode lifted interactivity by 2.2x to 2.4x at one to eight concurrent users, while DSpark added about a third more interactivity at one to four users and 13% at eight when composed with mono decode. AMD said DSpark without mono decode is the fallback for setups where mono cannot run.
For ATOM, AMD said it applies per-channel FP8 attention and a fused shared expert at startup, and captures the DSpark draft pass with CUDA graphs. The company said the draft accepts about half of its seven speculative tokens on real text. AMD also said the MI350 implementation uses 4-way tensor parallelism per replica, allowing two replicas on an 8-GPU node, and that both engines support the release checkpoint and the Quark MXFP4 export, though ATOM serves the Quark export in its reported recipe.
The article included benchmark tables from October 2026 showing that, on AMD’s test setups, vLLM and ATOM produced closely grouped results across GSM8K, AIME 2025, AIME 2026 and MMMU-Pro, with the best reported accuracy varying by engine and checkpoint. AMD also noted limits to the current mono-decode implementation: the 16-token step cap means mono plus DSpark is only usable up to four concurrent requests in the cited configuration, while wider decode steps fall back to the standard path. AMD said wider mono steps and ATOM mono-decode support for MiniMax-M3.1 are next on its development list, but did not give a date for that work.
