Xiaomi has released and open-sourced its MiMo-V2.6 model family, putting a major emphasis on reinforcement learning rather than simply scaling pretraining.

The release includes MiMo-V2.6-Pro and MiMo-V2.6-Flash, along with training resources and supporting infrastructure.

What makes MiMo-V2.6 different?

Xiaomi focused heavily on large-scale reinforcement learning.

The company describes the approach as a path toward recursive self-improvement, where models learn through repeated interaction with tasks and feedback.

The six-day training runs reportedly cost approximately $2.62 million for Pro and $850,000 for Flash. Xiaomi says the runs processed roughly 750,000 trajectories and produced significant improvements on its evaluation tasks.

Why does reinforcement learning matter?

Traditional model training relies heavily on large datasets.

Reinforcement learning introduces another mechanism.

The model attempts a task, receives feedback and adjusts its behavior.

This is particularly useful for problems where success can be verified, such as:

Coding

Mathematics

Cybersecurity

Tool use

Agent workflows

MiMo-V2.6 uses more than 7,000 task environments covering software engineering, vulnerability reproduction, knowledge-intensive work and web development.

The cost question

Large-scale RL is not cheap.

The MiMo-V2.6 runs demonstrate that improving model intelligence through extended interaction requires significant compute.

That creates an important industry trade-off.

More training compute can improve capability, but companies need to determine whether the resulting improvement justifies the additional cost.

Why the release matters

Xiaomi has also open-sourced much of its training infrastructure and model resources.

That gives researchers an opportunity to study how large-scale agentic reinforcement learning works in practice.

MiMo-V2.6 therefore represents more than another open model release.

It is an example of where AI development is heading: increasingly sophisticated training environments, stronger reinforcement learning and more emphasis on models that can actually complete complex tasks.