MiniMax M3.1 Flash: A Cheap Open-Weight Model for Coding Agents

Self-reported benchmarks (model card)

Order: M3.1 Flash / DeepSeek V4.1 Flash-1017 / Claude Opus 5, as published in MiniMax’s table; a dash means MiniMax lists no score. Harnesses vary by row: DeepSWE uses mini-SWE-agent at 512K context, Terminal Bench 4.0 uses Codex, CyberGym uses Claude Code at 1M.

• • Automation-Bench v1.0.6: 62.8 / 60.5 / 50.3
• • BankerToolBench: 76.2 / — / 76.2
• • DeepSWE v1.1: 73.7 / 77.0 / 74.0
• • Terminal Bench 4.0: 38.1 / 42.3 / 51.8
• • NL2Repo-Bench: 61.2 / 68.2 / 75.3
• • SWE Refactor Bench: 23.0 / — / 31.0
• • OSWorld 2.0: 61.7 / — / 70.6
• • CyberGym: 82.2 / — / —
• • HLE (text-only subset, no tools): 43.4 / 42.6 / 55.6

Python: Anthropic SDK (recommended)

MiniMax’s sample sends effort max; I set high to show the knob. Thinking comes back as its own content block.

# pip install anthropic
import os
import anthropic

client = anthropic.Anthropic(
    base_url="https://api.minimax.io/anthropic",
    api_key=os.environ["MINIMAX_API_KEY"],
)

message = client.messages.create(
    model="MiniMax-M3.1-Flash",
    output_config={"effort": "high"},  # low | medium | high | xhigh | max (default: max)
    max_tokens=4096,
    messages=[{"role": "user", "content": "Hi, how are you?"}],
)

for block in message.content:
    if block.type == "thinking":
        print(f"Thinking:\n{block.thinking}\n")
    elif block.type == "text":
        print(f"Text:\n{block.text}\n")

Python: OpenAI SDK

Swap base_url and model in an existing OpenAI client. On this protocol the thinking arrives in reasoning_content, and content holds only the answer.

# pip install openai
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.minimax.io/v1",
    api_key=os.environ["MINIMAX_API_KEY"],
)

response = client.chat.completions.create(
    model="MiniMax-M3.1-Flash",
    reasoning_effort="high",
    messages=[{"role": "user", "content": "Hi, how are you?"}],
)

print(response.choices[0].message.content)

Claude Code: ~/.claude/settings.json (verbatim from MiniMax’s docs)

Then run claude in your project and check that /status shows api.minimax.io/anthropic and /model shows MiniMax-M3.1-Flash.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.minimax.io/anthropic",
    "ANTHROPIC_AUTH_TOKEN": "<MINIMAX_API_KEY>",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "524288",
    "CLAUDE_CODE_EFFORT_LEVEL": "max",
    "ANTHROPIC_MODEL": "MiniMax-M3.1-Flash[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "MiniMax-M3.1-Flash[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "MiniMax-M3.1-Flash[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "MiniMax-M3.1-Flash[1m]"
  }
}

OpenCode: ~/.config/opencode/opencode.json (only if the model isn’t listed)

First run opencode auth login, pick “MiniMax Token Plan (minimax.io)” and paste your Subscription Key. Add this block only if your OpenCode version doesn’t list MiniMax-M3.1-Flash yet.

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "minimax-coding-plan": {
      "models": {
        "MiniMax-M3.1-Flash": {
          "name": "MiniMax-M3.1-Flash",
          "limit": { "context": 524288 }
        }
      }
    }
  }
}

vLLM on one 8×H100/H200 node (recipe command, verbatim)

The Hopper variant from the recipe: FP8 KV cache, NVFP4 experts on the Marlin kernel, 256K max context, tool calling on, video off.

hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash

vllm serve MiniMaxAI/MiniMax-M3.1-Flash \
  --tensor-parallel-size 8 \
  --reasoning-parser minimax_m3 \
  --block-size 128 \
  --kv-cache-dtype fp8 \
  --attention-config '{"indexer_kv_dtype": "fp8"}' \
  --compilation-config '{"cudagraph_mode": "FULL_AND_PIECEWISE"}' \
  --max-model-len 262144 \
  --max-num-seqs 80 \
  --limit-mm-per-prompt '{"video": 0}' \
  --gpu-memory-utilization 0.9 \
  --no-enable-log-requests \
  --disable-hybrid-kv-cache-manager \
  --moe-backend marlin \
  --enable-auto-tool-choice \
  --tool-call-parser minimax_m3

SGLang: the dev images from the cookbook

Support isn’t merged into SGLang main yet. Pull the dev image, then generate the launch command with the cookbook’s hardware picker.

docker pull lmsysorg/sglang:dev-minimax-m31-flash
# AMD MI350X / MI355X (ROCm 7.2):
docker pull lmsysorg/sglang:dev-minimax-m31-flash-mi35x