Xiaomi throws its hat into the open-source AI race
The newly launched AI model, MiMo-V2-Flash, particularly excels in reasoning, coding and agentic scenarios, said the company
This week, Chinese smartphone and electric vehicle maker Xiaomi has unveiled MiMo-V2-Flash, a new open-source artificial intelligence model aimed squarely at the world's leading systems.
The model is now available worldwide through Xiaomi's MiMo Studio, Hugging Face and its API platform. The company says it performs strongly in reasoning, coding and agentic tasks.
It is also designed to work as a general assistant for daily use. The move places Xiaomi in direct competition with models from DeepSeek, Moonshot AI, Anthropic and OpenAI.
MiMo-V2-Flash is part of Xiaomi's broader push into advanced AI. Luo Fuli, a rising figure in China's AI scene who recently joined Xiaomi after working at DeepSeek, described the release as the second step on the company's path towards artificial general intelligence. That term refers to systems that could one day match or exceed human thinking.
Speed and cost are central to Xiaomi's pitch — the company says the model processes up to 150 tokens per second, which makes it one of the most cost-effective high-performance models available.
On technical benchmarks, Xiaomi reports results close to Moonshot AI's Kimi K2 Thinking and DeepSeek V3.2 Thinking. In some long-context tests, it says MiMo-V2-Flash performs better.
On software engineering tasks, the model scored 73.4 per cent on SWE-Bench Verified, beating other open-source rivals and nearing the performance of OpenAI's GPT-5-High. Xiaomi also says its coding ability matches Claude 4.5 Sonnet at a far lower price.
The model uses a Mixture-of-Experts design and has 309 billion parameters. This allows it to handle complex tasks more efficiently. Xiaomi wants the AI model to run across phones, tablets and electric vehicles.
