Xiaomi Launches MiMo-V2-Flash Open-Weight AI Model to Challenge DeepSeek
A new open-weight AI model called MiMo-V2-Flash has been introduced to address complex reasoning, coding, and agentic AI tasks. At the same time, it is positioned as a general-purpose assistant for everyday use. According to shared details, MiMo-V2-Flash delivers inference speeds of up to 150 tokens per second. Additionally, it operates at a reported cost of $0.1 per million input tokens and $0.3 per million output tokens. The model includes 309 billion parameters, which typically indicate higher processing capability.
Moreover, the model is now publicly available through a developer portal, Hugging Face, and an API platform. As a result, it becomes the latest addition to the MiMo family of models. While the launch signals a move beyond hardware, it also places the company in direct competition with established AI players such as DeepSeek, Anthropic, and OpenAI.
Architecture, efficiency, and benchmarks
MiMo-V2-Flash uses a Mixture-of-Experts architecture to divide large neural networks. As a result, the design helps balance performance with efficiency. Furthermore, the model lowers the cost of processing long prompts by limiting how much prior context must be re-evaluated. This method supports scalability without sharply increasing compute requirements.
In benchmark testing, MiMo-V2-Flash performed on par with Moonshot AI’s Kimi K2 Thinking and DeepSeek V3.2 across most reasoning tasks. Additionally, it exceeded Kimi K2 in long-context evaluations. The model also outperformed all open AI model rivals on SWE-Bench Verified, where it achieved a score of 73.4 per cent. Meanwhile, its coding performance was reported to match that of Anthropic’s Claude 4.5 Sonnet, despite being developed at a significantly lower cost.
Role in a broader AI roadmap
The release aligns with efforts to roll out AI agent-driven features across smartphones, tablets, and electric vehicles. At the same time, progress in large AI models and applications has accelerated faster than expected. Moreover, deeper integration of AI with the physical world is increasingly seen as a key direction for future technology development.
