Xiaomi Develops MiMo-V2.5 AI Model, a New Competitor to Claude Opus and Gemini

Technology company Xiaomi has launched a new AI model called MiMo-V2.5. MiMo-V2.5 comes in two versions: the regular MiMo-V2.5 and MiMo-V2.5 Pro. Both are designed as open-weight AI models, meaning their weights are available to the public for independent use, but they are not fully transparent like open-source AI models. MiMo-V2.5 was trained using 48 trillion tokens and is designed as a native multimodal model that supports simultaneous processing of text, images, and videos. The regular MiMo-V2.5 features a total of 310 billion parameters (15 billion active), while the MiMo-V2.5 Pro has 1.02 trillion parameters (42 billion active). This AI model is claimed to excel in agentic AI as well as multimodal understanding. To prove the superiority claims of its latest AI model, Xiaomi has shared benchmark results for MiMo-V2.5 compared to other AI models such as DeepSeek-V4, Kimi K2.6, Claude Opus 4.6, Gemini 3.1 Pro, and MiMo-V2 Pro. Compared to several of these AI models, MiMo-V2.5 is claimed to outperform particularly in executing agentic AI tasks, as evidenced by the scores obtained in the company’s internal benchmarks. For example, in the MiMo Coding Bench benchmark, the standard MiMo V2.5 version is even claimed to match the performance of the Pro version, with comparison scores of 71.8 and 71.5, as shown in the following image. Meanwhile, in other benchmarks testing the AI model’s understanding of images and videos, MiMo-V2.5’s performance is on par with closed-source AI models like Claude Opus 4.6 and Gemini 3.1 Pro. In the VideoHolmes benchmark, MiMo-V2.5 achieved a score of 64.0, which is higher than Claude’s 59.5 and slightly below Gemini’s 64.2. Although available openly, running this model locally requires devices with high specifications, such as a Mac Studio with maximum configuration. This is because consumer GPUs are generally considered to lack sufficient VRAM capacity for MiMo-V2.5. For users eager to try it, they can access the AI studio platform or official API. The AI model can also be downloaded through AI platforms like Hugging Face, as compiled by KompasTekno from GSM Arena.