I just saw that Xiaomi's MiMo team open-sourced a new model, the V2.5 series.

What I find interesting is that they used the MIT license, which is very open. You can use it to develop commercial products, continue training it yourself, modify it freely—no one will bother you.
Let me also talk about the two models.
The Pro version is a pure text MoE with a total of 1.02 trillion parameters, but don’t panic; when actually running, only 42 billion are activated, so the hardware requirements aren’t that heavy. It’s mainly designed for agent tasks and coding.
Its score on ClawEval is roughly on par with GPT-5.4, but there’s an attractive data point: each task only consumes about 70k tokens, which is more than half less than other models.
This means that for the same work, your token bill can be significantly lower.
The other is a multimodal version with 310 billion parameters, activating 15 billion, capable of seeing, hearing, and reading images.
It’s equipped with dedicated visual and audio encoders.
Both models can handle around 1 million tokens of context at once, enough for long code projects or entire books.
They also launched a promotion: free 1 quadrillion token quota for 30 days.
Individuals, teams, and enterprises can apply; after that, they can use it on tools like Claude Code and Cursor.