Serve MiniMax M3 yourself for agentic coding (vLLM)

The Stack

Tested Against

Prerequisites

• A multi-GPU node (8x H200) or a rented GPU instance, to actually serve it

Steps

1. 1

   Write the serve command + agent endpoint, and validate them

   M3 support is not in a stable vLLM release yet, so use the dedicated image. --block-size 128 is mandatory (it matches the MSA sparse-attention cache; the default 16 fails to start). Any agent that takes a custom OpenAI base URL drives it. CI checks the serve command pins the model id + required flags and the endpoint config is well-formed.

   cat > serve.sh <<'SH'
   docker pull vllm/vllm-openai:minimax-m3 && vllm serve MiniMaxAI/MiniMax-M3 --tensor-parallel-size 8 --block-size 128 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice
   SH
   cat > endpoint.json <<'JSON'
   {
     "base_url": "http://localhost:8000/v1",
     "api_key": "EMPTY",
     "model": "MiniMaxAI/MiniMax-M3"
   }
   JSON
   node -e 'const fs=require("fs");const s=fs.readFileSync("serve.sh","utf8");const e=JSON.parse(fs.readFileSync("endpoint.json","utf8"));const need=["MiniMaxAI/MiniMax-M3","--block-size 128","--tool-call-parser minimax_m3","--reasoning-parser minimax_m3"];for(const t of need){if(!s.includes(t)){console.log("BAD: serve command missing "+t);process.exit(1)}}if(!String(e.base_url).includes("localhost:8000/v1")){console.log("BAD: endpoint base_url");process.exit(1)}if(e.api_key!=="EMPTY"){console.log("BAD: endpoint api_key should be EMPTY");process.exit(1)}if(e.model!=="MiniMaxAI/MiniMax-M3"){console.log("BAD: endpoint model id");process.exit(1)}console.log("config OK: serve pins MiniMaxAI/MiniMax-M3 with --block-size 128 + minimax_m3 parsers, agent endpoint well-formed")'

2. 2

   Point your agent at it (the model step, not checked by CI)

   In Aider: aider --model openai/MiniMaxAI/MiniMax-M3 --openai-api-base http://localhost:8000/v1 --openai-api-key EMPTY. In OpenCode/Cline/Kilo Code, add a custom OpenAI-compatible provider at the same base URL. The serving and the coding are fenced, because no green check stands up the model.