Xiaomi releases MiMo-V2.6 models and reinforcement-learning resources

THE ESSENTIALS

Xiaomi released Pro and Flash models plus reinforcement-learning research resources.
The Pro model card lists a one-million-token context and 1.02 trillion total parameters.
Published training costs describe the disclosed RL run, not the full cost of building the models.

A release centered on how the models were trained

Xiaomi released the MiMo-V2.6 Pro and Flash models on September 22, pairing the checkpoints with reinforcement-learning materials intended to let other researchers examine the training process. The announcement follows a six-day public training demonstration and puts the emphasis on improving agents through tasks, feedback and repeated attempts.

Xiaomi says the release includes a technical report, more than 7,000 task environments, training code and a distilled Qwen-based nine-billion-parameter model. The environments cover software engineering, knowledge work, web development and vulnerability reproduction. Publishing those components gives researchers more to investigate than a benchmark chart or a hosted chat interface alone.

What Xiaomi disclosed

The company reported approximately 750,000 trajectories across the disclosed run, with training costs of about $850,000 for Flash and $2.62 million for Pro. These figures concern the specified reinforcement-learning exercise. They should not be interpreted as the complete cost of pretraining, data preparation, research staff or the infrastructure used to develop the entire model family.

The Pro-RL model card describes a sparse mixture-of-experts architecture with 1.02 trillion total parameters and 42 billion activated parameters, a one-million-token context, and text, image, audio and video inputs. Activated parameters describe the subset used during a processing step; the full checkpoint still has to be stored and served. An open checkpoint of this scale is therefore a research and infrastructure release, rather than a promise of ordinary laptop deployment.

Why the grading process matters

Xiaomi's model card describes training that mixes several task types and agent frameworks. It also describes grading that compares solutions within a group, so successful attempts can still receive different rewards according to their quality. A binary pass or fail may establish that a test suite ran successfully, but cannot by itself distinguish a concise solution from a brittle one.

The team reports safeguards against reward hacking, including verifier cross-checks and adversarial screening. These details are relevant because an agent can learn to exploit weaknesses in a grader rather than complete the underlying task. The published resources make that claim more inspectable, although this article does not independently reproduce the training results.

The measured change during the public run

Xiaomi's disclosed DeepSWE v1.1 results give the six-day exercise a measurable endpoint. Flash rose from 48.8 to 65.7, while Pro moved from 58.4 to 72.6. These are roughly 17-point and 14-point gains on that software-engineering evaluation, rather than the percentage changes in the separate training-task pass rates. The comparison concerns the checkpoints before and after the specified run.

The training announcement also identifies 1,568 samples per update and approximately 3.5 billion to 3.7 billion tokens per training step, using asynchronous execution. Xiaomi says it froze the mixture-of-experts router to control changes in expert load. Together with the public task environments, those settings give researchers concrete variables to examine when testing whether the gains transfer to another training setup.

Access and the limits of the comparisons

The models are available through Xiaomi's platform, and the announcement says standard API pricing remains aligned with the V2.5 series. The released Pro-RL card carries an MIT license. Xiaomi also published changes to MiMo Code for the new family, including multimodal file reading and more controlled scheduling of tool calls.

Xiaomi's reported benchmark improvements support its case for larger reinforcement-learning runs, but the precise model checkpoint, agent framework and evaluation settings remain part of each result. They are not a universal guarantee that either model will complete every software or office task more cheaply or accurately.