Xiaomi’s MiMo-V2-Flash Clears the AI Hurdle: A Fierce Competitor in the Field

The Chinese electronics manufacturer Xiaomi has rolled out to the public a new neural network, MiMo-V2-Flash, which, according to the company, is poised to compete with the latest systems from DeepSeek, Moonshot AI, Anthropic, and OpenAI. MiMo-V2-Flash is available to developers on platforms such as MiMo Studio, Hugging Face, and via API. Xiaomi claims that the model “excels in logical reasoning tasks, programming, and agent creation, and also serves as a great all-purpose assistant for everyday tasks.” Source: Xiaomi

Luo Fuli, an artificial intelligence specialist formerly with DeepSeek and recently joined the MiMo team at Xiaomi, remarked on the social network X that MiMo-V2-Flash is “the second step towards creating artificial general intelligence (AGI).” AGI is a hypothetical type of AI that reaches or surpasses human cognitive capabilities.

Xiaomi claims that MiMo-V2-Flash is designed for maximum efficiency. The model offers a high data processing speed of 150 tokens per second, with a usage cost of $0.1 per million input tokens and $0.3 per million output tokens. According to the company, this makes it one of the most cost-effective models with high performance. The MiMo-V2-Flash is a Mixture of Experts model with a total of 309 billion parameters and 15 billion active parameters.

Xiaomi states that MiMo-V2-Flash showcases performance comparable to Moonshot AI’s Kimi K2 Thinking and DeepSeek V3.2 Thinking across most logical reasoning benchmarks, while maintaining competitive general-purpose writing capabilities. In evaluations of long-context processing, Xiaomi’s new model surpasses Kimi K2 Thinking. In agent creation tasks, MiMo-V2-Flash scored 73.4% in SWE-Bench Verified, outstripping all open-source competitors and coming close to the performance of OpenAI’s GPT-5-High. The model also stands on par with Claude 4.5 Sonnet, reputedly the best model for programming in the industry, at a significantly lower cost.