MiMo-V2.6-Pro-RL

MiMo-V2.6-Pro-RL

Works with text, images, video and audio in one 1.02T-parameter model that has a 1M token context. The flagship of Xiaomi's MiMo-V2.6 series, trained with large-scale reinforcement learning for coding and agent work.

Made by XiaomiMiMo on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiMo-V2.6-Pro-RL is the lead checkpoint in Xiaomi's MiMo-V2.6 line. It is a sparse mixture-of-experts model with 1.02T parameters in total and 42B active at a time.

One model takes in text, images, video and audio, and its context reaches 1M tokens. That length is meant for large code repositories, tool traces and agent runs that span several sessions.

Who it suits: Teams with serious infrastructure who build coding assistants, long-running agents or multimodal tools, and researchers who study reinforcement learning at scale.

Industry | Who uses it | How they use it | Example
Software | Teams building coding agents | Feed whole repositories and tool traces into one context | Work through a large codebase over several sessions
Research | Reinforcement learning researchers | Study mixed RL and groupwise grading | Compare against the smaller Flash-RL sibling
Media | Teams working across modalities | Combine text, image, video and audio inputs | Ask one model about a clip and its transcript

Skip it if you have limited hardware, since the weights alone come to about 1,024 GB.

Where to use it:

• Windows, macOS or Linux with Python, to download the files
• Xiaomi MiMo API Platform, Xiaomi MiMo Studio and Xiaomi MiMo Desktop, which the card lists
• Large GPU servers for self-hosting

Use case | Who | What they get back
Long-horizon coding | Developers | A model with a 1M token context for repositories and tool traces
Agent runs | Agent builders | A flagship RL-trained model for multi-session tasks
Multimodal questions | Product teams | Answers that draw on text, image, video and audio

How to set it up:

• Python and pip to run the download commands
• Large storage, since the weights are about 1,024 GB in U8 by Hugging Face figures
• Serving software and hardware suited to a mixture-of-experts model of this size

1. Install the Hugging Face tools.
2. Download the model files.
3. Serve the model with your own stack, or try the hosted options the card lists.

pip install -U huggingface_hub
hf download XiaomiMiMo/MiMo-V2.6-Pro-RL

How to use it: The visible part of the card gives no run command, so use the serving notes on the model page. Once it runs, send text and optionally images, video or audio as input.

Capability | What it supports
Omnimodal input | Takes text, pictures, video and sound together in a single model
Long context | Up to 1M tokens
Sparse mixture of experts | 42B of 1.02T parameters active
Speculative decoding | A multi-token prediction decoder with 5 layers

Example: You give the model a large repository plus a failing test and ask for a fix. The card describes this kind of long coding and tool trace work as a target, so expect a text answer with code.

For newcomers and experts: If you are new to large models, treat this as a research-grade flagship that needs a server, not a laptop model. Try the Xiaomi hosted options first if you only want to see how it answers.

If you know the field, the card says a single RL run mixed coding, general agent, visual and security tasks together. It uses fully asynchronous GRPO with 1,568 prompts and 16 rollouts per step, groupwise agentic grading, and a later multi-teacher distillation stage.

Mixed RL across coding, agents, vision and security
Groupwise grading ranks passing solutions
Multi-teacher on-policy distillation
Flagship MiMo-V2.6-Pro-RL checkpoint

Good to know:

• The model has a smaller sibling, MiMo-V2.6-Flash-RL, listed on the same download table
• Hugging Face lists 10 quantized versions made by other people
• The shown part of the card has no benchmark numbers, since the evaluation section is cut off

Title: Encoder sizes in millions of parameters
Vision encoder: 681 million
Audio tokenizer: 308 million
Audio patch encoder: 127 million

• License: MIT
• Cost to run: none named, though self-hosting needs hardware on a huge scale and the card also lists Xiaomi hosted services
• Data it reaches: the text, images, video and audio you send it
• Not for: the source states no restrictions

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 1T
• Weights: about 1,024 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 24,207
• Likes: 818
• License: mit
• Languages: en, zh
• Quantized versions: 10
• Fine-tunes: 3
• Tags: transformers, safetensors, mimo_v2, text-generation, multimodal, vision-language, audio, agent, video-understanding, long-context, conversational, 8-bit

Links

• Model: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL
• Files: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download XiaomiMiMo/MiMo-V2.6-Pro-RL

Then follow the usage in its model card. Nothing in this kit runs by itself.