MiMo-V2.6 is an environment dataset for LLM agent reinforcement training released by XiaomiMiMo in 2026, designed to provide multi-domain, verifiable RL training environments for LLM agents.

The dataset covers five domains: Code, Cyber, General, Visual, and Music, corresponding respectively to task families such as software engineering, vulnerability reproduction, knowledge work, web development, and symbolic music composition. It employs verification methods including executable tests, rule checks, rubric-based judging, or visual scoring. The dataset also provides training environments in the form of Docker images, intended to be used with verl training code.

Dataset Composition

• Code: Targeted at software engineering tasks, using executable tests as verifiers.
• Cyber: Targeted at vulnerability reproduction tasks, using rule checks as verifiers.
• General: Targeted at knowledge work tasks, using rubric-based judging as verifiers.
• Visual: Targeted at web development tasks, using visual scoring as verifiers.
• Music: Targeted at symbolic music composition tasks, using rule checks as verifiers.
• Runtime Environment: The dataset is distributed via Docker images and is associated with verl training code.
• Update Notes: The Code image includes built-in anti-hack preprocessing, and it is recommended to keep the anti_hack_cleanup configuration set to true.