MiMo-V2.6-Distill-Qwen-9B

MiMo-V2.6-Distill-Qwen-9B

A 9B agentic model from Xiaomi MiMo, fine-tuned from Qwen3.5-9B for coding, agent tasks, visual coding and cybersecurity, released as a starting point for research.

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-Distill-Qwen-9B comes from Xiaomi MiMo and is built for agent work, at about 9B parameters. It starts from Qwen3.5-9B and adds supervised fine-tuning on data generated by MiMo. The covered areas are coding, general agent tasks, visual coding and cybersecurity.

Researchers can use the released checkpoint as a base for their own agent reinforcement learning work. It reads images and text and answers in text.

Who it suits: It suits researchers and developers who want a mid-size open model for agent and coding work, or a base for their own reinforcement learning experiments. You need hardware that can hold about 18.8 GB of BF16 weights.

Industry | Who uses it | How they use it | Example
Software | Developers | Test a local model on coding and terminal tasks | Fixing an issue in a code repository
Research | Reinforcement learning researchers | Start from the released checkpoint for agent training | Running their own training on top of it
Office automation | Tool builders | Try a model on general agent tasks | Driving a task through tool calls
Security | Defensive teams | Study a model trained with cybersecurity data | Reading and explaining a code issue

Skip it if you only want a hosted chat assistant and do not run models.

Where to use it:

• SGLang, a model server, in a recent build with Qwen3.5 support
• Any client that talks to an OpenAI-style endpoint
• Windows, macOS or Linux, with Python for the file download

Use case | Who | What they get back
Coding and terminal tasks | Developers | Model answers and actions for repository work
Visual coding | Front-end developers | Code written from a visual description
Agent research | Researchers | A starting checkpoint for reinforcement learning work

How to set it up: You need a few things in place first.

• Python and pip
• A recent SGLang build that supports Qwen3.5
• Memory for the weights, about 18.8 GB in BF16

1. Install a recent SGLang build with Qwen3.5 support.
2. Start the server with the serve command, which names the model to load. The checkpoint brings its tokenizer and the MiMo v2.6 chat template.
3. Send requests to the server from an OpenAI-style client, with thinking turned on.

sglang serve \
  --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B \
  --reasoning-parser mimo \
  --host 0.0.0.0 \
  --port 30000

To only get the repository files onto your computer, use Hugging Face's own steps.

pip install -U huggingface_hub
hf download XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B

How to use it: Once the server is up on port 30000, you send chat requests to it. The card's client call sets enable_thinking to true, so the reply carries a thinking part and an answer part. The reasoning parser option separates the two.

Example: You ask the running server "What is 15% of 240?" with thinking enabled and a limit of 2048 tokens. The reply comes back in two parts, the model's reasoning first and then its final answer.

For newcomers and experts: Start by running the serve command and sending one short question. That tells you whether your setup works before you try coding tasks. Keep the first request small.

The card reports gains over Qwen3.5-9B on most of its benchmarks. On SWE Pro the score goes from 32.0 to 44.6, on Terminal Bench 2.1 from 27.0 to 37.1, and on AutomationBench from 5.0 to 30.3. Some benchmarks are internal sets. The fine-tuning mix holds 77.4B tokens, of which 27.2B carry loss.

Title: Benchmark scores from the card, base model against this model
SWE Pro, Qwen3.5-9B: 32.0 points
SWE Pro, this model: 44.6 points
Terminal Bench 2.1, Qwen3.5-9B: 27.0 points
Terminal Bench 2.1, this model: 37.1 points
AutomationBench, Qwen3.5-9B: 5.0 points
AutomationBench, this model: 30.3 points

Download the model or name it in the serve command
Start the SGLang server
Send a chat request with thinking on
Read the thinking and the answer

Good to know:

• License: MIT
• Cost to run: none named, but it needs hardware for about 18.8 GB of BF16 weights
• Data it reaches: it downloads model files from Hugging Face and serves requests on the machine you run it on
• Not for: unlawful or malicious use, military use, malicious cyber activity such as malware or phishing, or high-risk actions run without human oversight

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 9.4B
• Weights: about 18.8 GB in BF16; about 4.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 21,525
• Likes: 703
• License: mit
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 67
• Fine-tunes: 31
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, mimo_v2, agentic, distillation, supervised-fine-tuning, code, tool-use, conversational

Links

• Model: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B
• Files: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B/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-Distill-Qwen-9B

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