Xiaomi Launches MiMo-V2-Flash
Xiaomi has introduced MiMo-V2-Flash, a sophisticated open-source intelligence model designed to deliver high-speed reasoning and efficient processing across its integrated device network.
On December 16, 2025, Xiaomi introduced its latest breakthrough in artificial intelligence, a model known as MiMo-V2-Flash. Unveiled during a major partner summit focused on the company's interconnected ecosystem, this release highlights a strategic move toward creating advanced digital intelligence for cars, homes, and mobile devices. Under the direction of Luo Fuli, the development team designed this system to bridge the gap between language processing and real-world application.
Innovative Design and High-Speed Performance
The engine behind this new model is a specialized architecture that manages a massive 309 billion total parameters. However, it is designed for extreme efficiency, activating only 15 billion parameters at any given moment. This selective activation allows the system to process information with the depth of a large model while maintaining the agility of a much smaller one.
One of the most impressive features is the generation speed, which reaches 150 tokens per second. This is made possible through a technique that predicts several words at once rather than one by one. By generating and verifying multiple tokens in parallel, the system effectively doubles the output speed found in previous generations.
Efficient Memory and Massive Context
To handle large amounts of data, the model uses a hybrid attention system. This design significantly cuts down on the memory required to store information during a conversation. As a result, the model can maintain a context window of 256,000 tokens. This capability allows it to stay coherent during very long interactions or when analyzing massive documents, making it a powerful tool for complex automation.
Benchmarking Against Industry Leaders
The performance of MiMo-V2-Flash places it at the top of the open-source market. In rigorous testing, it has proven to be a formidable opponent for both open and proprietary rivals.
In mathematical evaluations, the model achieved a 94.1% success rate, which is higher than the performance of DeepSeek-V3.2. It also showed strong results in scientific knowledge, standing on level ground with Google’s Gemini 3.0 Pro. When it comes to software engineering, Xiaomi’s model matched leading competitors while operating at a much higher speed. Compared to premium models like Claude 4.5 Sonnet, the Xiaomi system provides similar technical results at approximately two percent of the cost.
Broad Accessibility and Integration
Xiaomi has released the weights and code for the model under the MIT license, allowing the developer community to use and modify it freely. The pricing for the cloud-based version is also set at a very low rate, making high-tier intelligence accessible to independent creators and small businesses.
Beyond its technical specs, the model is intended to be the central intelligence for Xiaomi’s hardware, including the SU7 electric car and various smart home products. By enabling devices to communicate in a common, efficient language, Xiaomi aims to create a more seamless and autonomous user experience across its entire product lineup.
