GitHub radar
Xiaomi's MiMo V2.6 flagship: a 1-trillion agent model
Xiaomi released the most capable model in its MiMo V2.6 line — a clear jump on agent and coding tasks, but it only runs on server hardware.
MiMo-V2.6-Pro-RL is described by its authors as the flagship checkpoint of the MiMo V2.6 series from the XiaomiMiMo team (Xiaomi's AI division). It is the largest model in the line: 1.02 trillion parameters in total, with about 42 billion active for each request, an architecture called a mixture of experts. The model can act as an agent: taking step by step actions in a terminal, fixing bugs in real code, reading interface screenshots, and holding a context of one million tokens. Compared with the previous flagship of the line, V2.5 Pro, the jump is especially visible on agent tasks: the AutomationBench test for completing tasks on a computer on its own rose from 16 to 53.1 out of 100 points, and the DeepSWE test for actual code fixing rose from 19 to 71.9. Xiaomi already released two smaller versions of this line, Flash-RL and the smaller Distill-Qwen-9B, which can be installed on a regular computer. Pro-RL is the top, most capable version, and it is built for server hardware only.
Why a vibe-coder should care
I need to say this plainly right away: this is not a model you will install on your laptop, it needs a data center. But the jump on agent tasks is real, a model that used to finish computer tasks almost by accident now completes the job in more than half of the cases. It shows where agent models are heading in general, and it is useful to know this level even if you work with Claude Code or another cloud agent yourself. For hands-on use there is not much point, unless Xiaomi opens an API for it.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the model card at https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL and break down what it can newly do on agent and coding tasks — and check whether there is a way to try it through an API or cloud service, since it cannot be installed locally.
This is a server-only model — 1.02 trillion parameters, you cannot run it at home. Even a compressed version would take about 615 GB.
Open on Hugging Face▌ More finds