MiniMax-M3
serving recipe.

Floors are the minimum the vLLM recipe publishes across the whole replica, from vLLM's own runs. The control plane compares them with its own sizing from the model config.

• Tensor ParallelFits inside one node's interconnect. single_node_tp
• Tensor + Expert ParallelFits inside one node's interconnect. single_node_tep
• Data + Expert ParallelFits inside one node's interconnect. single_node_dep
• Multi-Node Tensor ParallelSpans nodes. multi_node_tp
• Multi-Node TP + Pipeline ParallelSpans nodes. multi_node_tp_pp
• Multi-Node Tensor + Expert ParallelSpans nodes. multi_node_tep
• Multi-Node Data + Expert ParallelSpans nodes. multi_node_dep
• Prefill/Decode DisaggregationSpans nodes. pd_cluster

• B200 · 8 GPUs
• B300 · 8 GPUs
• H200 · 8 GPUs
• MI300X · 8 GPUs
• MI355X · 8 GPUs

Wingsmith runs models through NVIDIA Dynamo, so it serves on NVIDIA GPUs. Other vendors are listed as the recipe states them.

vllm serve MiniMaxAI/MiniMax-M3 \
  --block-size 128 \
  --tool-call-parser minimax_m3 \
  --enable-auto-tool-choice \
  --reasoning-parser minimax_m3

Features

• Tool callingOn by defaultMiniMax M3 tool call parser with automatic tool choice
• ReasoningOn by defaultMiniMax M3 reasoning parser (supports `thinking` and `non-thinking` modes, switchable per request)
• Speculative decodingOpt-inEagle3 speculative decoding with an Inferact MiniMax-M3 draft head — accelerates decoding. Choose between the standard MHA head and the GQA head (16× smaller draft KV cache — better for long-context or high-batch deployments). Modes: MHA, GQA.
• Text-only modeOpt-inSkip loading the vision encoder for text-only workloads — frees VRAM for KV cache. Mutually exclusive with encoder_parallel.
• Thinking always onOpt-inPin `thinking_mode: enabled` server-side — every turn reasons before answering, including after tool results. Recommended for agentic workflows.
• Encoder data parallelOpt-inRun the vision encoder data-parallel instead of tensor-parallel — avoids TP comm overhead on the small encoder. Also enables the host-shared-memory multimodal processor cache and the per-brand encoder attention backend (FlashInfer on NVIDIA, AITER FlashAttention on AMD). Mutually exclusive with text_only.

1. Register the weights

   In the console, or through the API. Resolving a slug proposes an artifact and writes nothing.

   curl "http://localhost:8000/v1/artifacts/resolve?source=hf:MiniMaxAI/MiniMax-M3"

2. Read the draft

   The draft carries this recipe as upstream evidence: the tested flags, the precision and memory floor, the parallelism layouts, and the hardware vLLM verified, with links back to the recipe. Nothing from it is applied without you, and the draft warns when the plane's own sizing and the published floor differ by a quarter or more.

3. Create the workload

   Choose the engine, device shape and scale. The plane picks the clusters, and each cluster's operator renders the deployment as a Dynamo graph.

4. Give it an endpoint

   One model name for clients, routed by the gateway to deployments on every cluster serving it.

Run the control plane locally →

• How much GPU memory does MiniMax-M3 need?The vLLM recipe publishes a minimum of 257 GB across the whole replica for its NVFP4 variant; BF16 1025 GB, MXFP8 513 GB, MXFP4 257 GB. That is vLLM's own measurement, not a per-GPU figure.
• Which parallelism layouts does the MiniMax-M3 recipe support?Tensor Parallel, Tensor + Expert Parallel, Data + Expert Parallel, Multi-Node Tensor Parallel, Multi-Node TP + Pipeline Parallel, Multi-Node Tensor + Expert Parallel, Multi-Node Data + Expert Parallel and Prefill/Decode Disaggregation.
• Can Wingsmith serve MiniMax-M3?Wingsmith serves models through NVIDIA Dynamo on clusters with NVIDIA GPUs. Register MiniMaxAI/MiniMax-M3 and the plane attaches this vLLM recipe to the draft as evidence: its tested flags, precision and memory floor, for you to review rather than applied on your behalf.

From the vLLM recipe for MiniMaxAI/MiniMax-M3 ↗ (snapshot at 4bcc189), Apache 2.0.