Xiaomi distills its MiMo agent into a 9B checkpoint

Xiaomi’s MiMo team has released MiMo-V2.6-Distill-Qwen-9B, a supervised fine-tune of Qwen3.5-9B trained on trajectories generated by the larger MiMo-V2.6 stack. The team positions the checkpoint as an open baseline for research into coding agents, tool use, visual coding, cybersecurity, and agentic reinforcement learning.

The release lowers the hardware and training barrier for teams studying agents. Xiaomi reports substantial gains over the original 9B base on several multi-step benchmarks, while the model remains small enough for single-GPU deployment under suitable memory, context, and quantization settings.

Distillation targets the action loop

Agentic models operate inside a loop: they choose an action or tool call, read the resulting observation, and decide what to do next. MiMo-V2.6-Distill-Qwen-9B learns those patterns from trajectories produced by a larger teacher. An external agent harness still needs to expose tools, execute calls, return observations, and enforce permissions.

• Base model: Qwen3.5-9B
• Training method: supervised fine-tuning on MiMo-generated data
• Inputs and outputs: images and text in, text out
• Precision: BF16 weights
• License: MIT
• Thinking mode: configurable through chat-template arguments
• Serving options: SGLang, vLLM, Transformers, and Docker Model Runner

Automation and cyber scores move most

Xiaomi reports improvements on all seven listed benchmarks, with the largest gains appearing in automation, cybersecurity, job tasks, and repository-level software engineering.