Use MiMo-V2.6-Pro from OpenRouter or Command Code, no GPUs

The self-hosting setup for MiMo-V2.6-Pro needs a multi-GPU server. This is the other route: rent it per token, through OpenRouter or through Command Code.

Where this comes from: OpenRouter's public model list and Command Code's docs, checked on 4 October 2026, plus two runs I made that day, one OpenRouter call and one Command Code run, described under their routes. The bring-your-own-key section follows Command Code's docs. Two runs are data points, not a benchmark.

What OpenRouter lists

Id	Context	Input per million	Output per million
xiaomi/mimo-v2.6-pro	1,050,000	$0.435	$0.87
xiaomi/mimo-v2.6-flash	1,050,000	$0.14	$0.28
xiaomi/mimo-v2.6-pro-ultraspeed	1,048,576	$4.35	$8.70
xiaomi/mimo-v2.5-pro	1,050,000	$0.435	$0.87

The Pro listing marks tools, reasoning and structured outputs as supported. Four providers serve it (DeepInfra, Novita, GMICloud and Xiaomi), with output caps from 131,072 to 945,000 tokens, so check the provider you land on. The ultraspeed listing costs ten times as much. OpenRouter describes it as a speed edition built from the same checkpoint.

Call it from OpenRouter

You need an OpenRouter account, some credit and an API key. Put the key in an environment variable without typing it into your shell history, and never in a file you commit.

printf 'OpenRouter key: '; read -rs OPENROUTER_API_KEY; echo; export OPENROUTER_API_KEY

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "xiaomi/mimo-v2.6-pro",
    "temperature": 1.0,
    "top_p": 0.95,
    "max_tokens": 400,
    "messages": [{"role": "user", "content": "In one sentence, what is a mixture of experts model?"}]
  }'

temperature 1.0 and top_p 0.95 are the values the model card recommends, and the same ones OpenRouter's listing shows as the defaults, so you can leave both out. max_tokens is the cap my call used. The listing supports reasoning, which can add output tokens.

What my one call did, with max_tokens capped at 400 and the prompt above: it returned 200 in 5.5 seconds, served by GMICloud, and the answer was one sentence. It used 19 prompt tokens and 87 output tokens, 21 of those reasoning, and cost $0.000084, which matches the listed rates. That is a single call, not a speed test.

Command Code, built in

Command Code's docs list xiaomi/mimo-v2.6-pro ("flagship multimodal agentic coding with 1M context"), and its 21 September announcement gives the command:

cmd --model xiaomi/mimo-v2.6-pro

It needs a current version. My copy was 0.52.1, and its --list-models had no V2.6 at all. The install is npm i -g command-code@latest on Node.js 22 or newer, and the update command is cmd update. cmd update failed for me, and printed the npm line to run by hand. That hit a permissions error too, because npm's default folder here needed root, so I added --prefix pointing at a folder I own: npm i -g --prefix "$HOME/.npm-global" [email protected]. On 1.66.0 the list showed xiaomi/mimo-v2.6-pro, and I ran the same one-sentence prompt with cmd -p "..." -m xiaomi/mimo-v2.6-pro. It answered in about 5.7 seconds of wall time. This mode prints no token counts, so I have no cost for it. The announcement lists Pro at $0.43 per million input tokens and $0.87 output, on its Go plan and above.

Or bring your OpenRouter key into Command Code

Use this if you'd rather OpenRouter bills you, or you want a model Command Code doesn't list. Command Code's bring-your-own-key docs name OpenRouter as a supported provider. Create ~/.commandcode/providers.json, or if the file already exists, add the openrouter entry under provider instead of replacing it. This is the docs' own OpenRouter example with the model id swapped and a context window added:

{
  "provider": {
    "openrouter": {
      "name": "OpenRouter",
      "baseURL": "https://openrouter.ai/api/v1",
      "apiKey": "$OPENROUTER_API_KEY",
      "models": {
        "xiaomi/mimo-v2.6-pro": { "contextWindow": 1000000 }
      }
    }
  }
}

The docs say {} per model id is enough, and that the context window defaults to 200K if you leave it out, so set it if you want the long context. Keep the $OPENROUTER_API_KEY reference as written, because the docs say a raw secret pasted into the file is refused. Export the variable in your shell, start cmd, type /model, and pick the model from the OpenRouter section. The docs describe picking it with /model.

The docs say requests go straight from your machine to your endpoint with your own keys, so OpenRouter should bill you. Check its activity page after the first call to be sure.

Where it will disappoint you

These are two runs, not benchmarks. The benchmark table on the self-hosting page is Xiaomi's own, so check it against your own work before you trust it.

Watch the bill on the first day. A full 1M token prompt would cost $0.435 at the listed rate, and a coding agent sends its history again every turn.

Prices change. These are from OpenRouter's public list and Command Code's announcement, read on 4 October 2026.

Source

OpenRouter's public model list, openrouter.ai/api/v1/models, its endpoints list for the model and the MiMo-V2.6-Pro page. Command Code: MiMo V2.6 announcement, models, bring your own key and the quickstart.

Where the longer version lives

The Rabbit Hole is the community where I post the setups, the ones that worked and the ones that didn't. Tell me what you'd point it at and I'll say which route fits. There's a step by step course for this route there too: Use MiMo-V2.6-Pro without a GPU. Joining is free, the course is open to every member, and Premium ($5 a month) adds weekly office hours. The other open courses are on the courses page. skool.com/the-rabbit-hole-3206