GitHub radar
Xiaomi's compact agent model now runs on your laptop
Xiaomi released a compact version of its new MiMo V2.6 agent model. Unlike the 309-billion-parameter flagship, this one runs on a regular laptop.
MiMo-V2.6-Distill-Qwen-9B is what the authors call - in their own words - "an agentic model with 9 billion parameters, fine-tuned on data generated by MiMo". The fine-tuning was done by the XiaomiMiMo team (Xiaomi's AI division). It's a compact version of the MiMo V2.6 line - the flagship of that line has 309 billion parameters and you can't run it at home. The model does four things: it writes and fixes code, acts step by step through tools like the terminal, reads interface screenshots, and solves cybersecurity tasks. On AutomationBench (a test for autonomous runs on a computer) the score went from 5 to 30.3 out of 100 compared to the regular Qwen3.5-9B of the same size. On SWE Pro (a test for fixing bugs in real code) it went from 32 to 44.6 out of 100. The authors themselves call this release not a finished product but a working snapshot - a starting point for further training of agent skills.
Why a vibe-coder should care
The MIT license and local run mean the model can be installed for free and without anyone else's keys. On quality I wouldn't compare it to Claude Code - it's a small model, and the authors themselves call the release an early version for further training. For playing with agent scenarios, for simple repetitive coding tasks, or when there is nothing to pay for - it's a workable option.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Xiaomi's MiMo-V2.6-Distill-Qwen-9B model locally for me: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B — pick a compressed version that fits my laptop, run it, and show me how to give it an agent task such as fixing code.
A regular laptop with 16 GB of RAM or more — the compressed version of the model takes about 6 GB.
Open on Hugging Face▌ More finds