Xiaomi introduced its new open-based artificial intelligence model, MiMo-V2.5, to the technology world. The company states that this new model makes significant progress in both agent capabilities and multi-modal understanding.

Xiaomi shared various test results comparing the MiMo-V2.5 model with current models such as DeepSeek-V4, Kimi K2.6, Claude Opus 4.6 and Gemini 3.1 Pro. The company claims that the MiMo-V2.5 delivers best-in-class performance in its self-developed intermediary tasks testing.

MiMo-V2.5 Specifications and Performance

In MiMo Coding Bench tests, the smaller V2.5 model manages to achieve the performance of the larger V2.5-Pro ​​model at half the cost. In other tests measuring image and video understanding abilities, V2.5 is at the same level as closed source models.

The model, trained with a total of 48 trillion tokens, has a native multi-modal structure that supports text, image and video data. Xiaomi has released two different versions: MiMo-V2.5 with 310 billion total parameters and MiMo-V2.5-Pro ​​with 1.02 trillion total parameters. MiMo-V2.5The series offers broad context support of 1 million tokens.

Access and Hardware Requirements

Users can download the model from Hugging Face and run it on their own systems. However, this process requires very powerful hardware; Because even the VRAM capacity of consumer-grade graphics cards such as Nvidia RTX 5090 is not enough to run this model locally.

For those who want to try the model, Xiaomi makes available the AI ​​Studio platform or official API service. Users who want to perform local installation must have high-equipped systems such as Mac Studio.

What do you think about this new generation artificial intelligence model of Xiaomi?