Xiaomi has launched its new open-source AI models, the MiMo-V2.5 and V2.5-Pro, which are tailored for efficient execution of agentic “claw” tasks, crucial for automating functions like content creation and scheduling through third-party applications. Released under the MIT License, these models allow enterprises and developers to seamlessly modify and integrate their capabilities into commercial applications without restrictive usage policies. The V2.5-Pro model is particularly notable for achieving a 63.8% success rate while utilizing significantly fewer tokens compared to leading competitors like Anthropic and OpenAI, thus providing a cost-effective solution in an era of rising usage-based billing for AI services. Additionally, the launch has garnered immediate support from the AI ecosystem, with integration into high-throughput inference engines such as SGLang and vLLM, enabling efficient performance across various computational platforms.

Xiaomi: Xiaomi is a Beijing-based technology company specializing in smartphones, electric vehicles, and a connected smart device ecosystem under HyperOS. It has pivoted into AI development, releasing open-source large language models through its MiMo project. In this news, Xiaomi launched MiMo-V2.5 and MiMo-V2.5-Pro, positioning them as efficient options for agentic tasks amid industry shifts to usage-based pricing.
Fuli Luo: Fuli Luo is the project lead for Xiaomi’s MiMo AI initiative and a former core contributor to the DeepSeek team with prior experience at Alibaba. She built Xiaomi’s large model division from scratch, emphasizing practical problem-solving in AI development. In this context, she promoted the MiMo-V2.5 series on X, announcing developer incentives to accelerate agentic building.
MiMo-V2.5: MiMo-V2.5 is Xiaomi’s open-source multimodal large language model supporting native text, vision, audio, and video processing with strong agentic capabilities. Released under the permissive MIT License, it balances omnimodal perception and task execution for developers and enterprises. The model is highlighted in the news for its efficiency in agentic claw tasks, enabling cost-effective deployments.
MiMo-V2.5-Pro: MiMo-V2.5-Pro is Xiaomi’s advanced open-source large language model specialized for agentic workflows, long-horizon reasoning, and complex software engineering. It features enhanced coherence over extended tool calls and is licensed under MIT for commercial production use. In the news, it leads in benchmarks for claw tasks, demonstrating prowess in autonomous projects like compiler building and circuit optimization.

Licensing: Models released under MIT License for unrestricted commercial deployment, fine-tuning, and derivative releases.
Agentic Focus: Specialized for claw tasks in systems like OpenClaw, sustaining coherence over extensive tool calls.
Ecosystem Support: Immediate Day-0 integration with SGLang, vLLM inference engines, and hardware from AMD and AWS.