mimo-v2.6-pro open weights and deployment

Status: Available. Xiaomi published the MiMo-V2.6-Pro-RL weights (about 573 GB of FP8 safetensors stored in MXFP4), which need custom code, under the MIT License declared in the model card metadata; the repository has no separate LICENSE file. Xiaomi's pages do not say in so many words that this checkpoint is the one behind the mimo-v2.6-pro API.

Agent deployment prompt

You are my deployment agent. Deploy the official XiaomiMiMo/MiMo-V2.6-Pro-RL weights as a private, OpenAI-compatible service on this host or cluster. Use the publisher's model card as the source of truth: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL.

Preflight
- Inspect the OS, GPU model and count on every node, the interconnect between nodes, free VRAM, driver and CUDA versions, free disk space, and available ports. The weights are about 573 GB of FP8 safetensors stored in MXFP4. The model card's SGLang example spans two nodes with tensor parallel 16 and its vLLM example uses tensor parallel 8, but it states no minimum GPU; confirm the GPUs support these weight formats and estimate whether the official weights and the requested context fit with safe headroom.
- The repository ships custom model code that needs trust-remote-code. Pin the exact revision you download, list the Python files it would execute, and ask me to approve them before enabling trust-remote-code.
- If the hardware cannot safely run the official weights, stop before installing or launching anything. Explain the limiting resource and propose a separately reviewed alternative; do not substitute a community quantization automatically.

Deployment
- If preflight passes and I have approved the remote code, create an isolated environment and follow the model card's SGLang (its recommended runtime) or vLLM launch instructions, including its reasoning-parser and tool-call-parser settings. The model card points to the MiMo-V2.5 SGLang cookbook and vLLM recipe and image; use the versions it names rather than untested ones. Generate an API key and store it in a protected environment variable. Bind to 127.0.0.1 first; do not expose the endpoint publicly.
- Keep the publisher's sampling defaults (temperature 1.0, top_p 0.95). Start with a context length the preflight supports; the model supports up to 1,048,576 tokens.

Verification and handoff
- Send a real request to /v1/chat/completions and confirm that the model responds. Report the exact launch command, installed versions, model revision, local endpoint, test result, measured GPU memory and log location. Identify any reverse-proxy or firewall changes that still need my approval. Never print the API key or other secrets.
License
MIT (declared in the model card metadata; no LICENSE file in the repository)
Hardware note
Not published by Xiaomi beyond the model card's launch examples (SGLang across two nodes with tensor parallel 16, or vLLM with tensor parallel 8); the Agent recipe performs a hardware preflight.
Recommended runtime
SGLang (recommended in the model card) or vLLM
Verified date
2026-09-24

Official repository

Published model card

Parameter count
1.02T total, 42B active (MoE), per the model card and technical report
Context window
1M tokens (1,048,576)
Maximum output
128K tokens (131,072, including reasoning tokens)
Input/output modalities
Text, image, video and audio input; text output

Official sources