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

A 9B agentic model Xiaomi ships as a shared starting point for open research on agentic RL, alongside the MiMo-V2.6 weights, roughly 7,000 released RL tasks with verifiers (about 3k software-engineering tasks with executable tests, 1k vulnerability-reproduction tasks with rule checks, 1k knowledge-work tasks with rubric judging, 2k web-development tasks with visual grading, plus about 1k music-generation tasks), and an end-to-end RL framework. It is a supervised fine-tune of Alibaba's Qwen3.5-9B on MiMo-generated trajectories: 77.4B tokens in total, 27.2B of them loss-bearing, split across code (23.2B), general agents (22.0B), visual (21.2B), and cybersecurity (11.0B). MIT license; thinking mode is enabled through the chat template.

The SFT checkpoint lifts Qwen3.5-9B on every reported evaluation: SWE-bench Pro 32.0 to 44.6, SWE-bench Verified 60.0 to 61.1, AutomationBench 5.0 to 30.3, Terminal-Bench 2.1 27.0 to 37.1, Toolathlon-Verified 25.9 to 35.2, JobBench 2.6 to 18.3. The report's point is what happens next: GRPO on the released environments from this checkpoint improves all 11 evaluations, including SWE-bench Verified to 66.2, Terminal-Bench 2.1 to 52.8, MiMo Cyber Bench (mini) 31.3 to 47.0, and MiMo Visual Coding (mini) 64.0 to 72.4, and multi-harness RL on coding improves all 21 dataset-harness pairs, including three held-out harnesses. Only the SFT checkpoint is released; the RL'd variants are baselines in the report.