샤오미 MiMo V2.5 Pro, AI 에이전트 경쟁에서 키미에 도전하며 글로벌 톱 5 진입 - kmjournal.net

The global AI race is shifting fast. What used to look like a tight contest led by U.S. tech giants is now opening up, with Chinese companies rapidly closing the gap. Xiaomi’s newly released MiMo V2.5 Pro sits right at the center of that shift, going head-to-head with Moonshot AI’s Kimi K2.6 and pushing into the global top tier. Xiaomi Enters the AI Agent Spotlight Xiaomi unveiled MiMo V2.5 Pro and its base model, MiMo V2.5, on April 22. Both are designed for agent-style AI tasks, meaning they don’t just respond to prompts but can carry out complex, multi-step work on their own. The focus is clear. These models are built to handle long-running tasks like software development, system design, and iterative problem-solving. Xiaomi is currently running a public beta and plans to release the models as open source in the near future. A Rapid Climb Into the Global Top 5 What stands out most is how quickly MiMo V2.5 Pro has climbed the rankings. According to Artificial Analysis, the model scored 54 points on the intelligence index, placing it fifth globally. Just a month ago, its predecessor sat at eighth place. That’s a three-step jump in a very short time. At the top of the chart are models from OpenAI, Google, and Anthropic, along with Moonshot AI’s Kimi K2.6. In terms of raw score, MiMo V2.5 Pro is now effectively tied with Kimi. This is more than a ranking shift. It shows that the AI landscape is no longer dominated by a single region. Instead, it’s becoming more distributed, with China forming its own competitive axis. Built for Long-Horizon Tasks MiMo V2.5 Pro is designed around one key idea: staying on track over long workflows. It can repeatedly perform actions like web searches, code execution, file handling, and API calls while maintaining the original objective. Even in tasks that require over 1,000 tool calls, the model keeps its context and consistency intact. A notable feature is its “harness awareness,” which allows the model to manage its own execution environment. It can adjust memory usage and optimize workflows on the fly. This moves it beyond simple instruction-following into active task management. Real-World Performance Backed by Results The model’s capabilities are not just theoretical. In a benchmark based on a Peking University compiler assignment, MiMo V2.5 Pro implemented a full SysY compiler in Rust. The task typically takes weeks, but the model completed it in just 4.3 hours and achieved a perfect score. In another test, it built a video editing application from scratch, generating over 8,000 lines of code and working autonomously for more than 11 hours. The process included design, debugging, and validation. It also showed strong performance in hardware-related tasks. In analog circuit design, the model interacted with simulation tools and iterated until it met all key performance metrics. This is the kind of work that usually takes experienced engineers several days. Lower Costs, Higher Efficiency Performance is only part of the story. Efficiency has improved as well. MiMo V2.5 Pro uses 40 to 60 percent fewer tokens compared to models with similar performance levels. That directly translates into lower operational costs. On the CloiVal benchmark, it achieved around 64 percent performance, outperforming competitors like Gemini 3.1 Pro in cost efficiency. For companies considering deployment, this makes a real difference. A Multimodal Model for Broader Use Alongside the Pro version, Xiaomi introduced MiMo V2.5 as a general-purpose model. It supports multimodal inputs, meaning it can understand and reason across text, images, and audio. It also handles up to one million tokens of context, allowing it to process very large datasets or long conversations. One key design choice is the integration of perception and action. Instead of separating understanding from execution, MiMo processes both within a single system. This sets it apart from many existing multimodal models. China’s AI Market Forms a Two-Player Race With this release, China’s AI landscape is starting to take shape around two major players: Moonshot AI’s Kimi and Xiaomi’s MiMo. Kimi has led in areas like long-horizon execution and multi-agent coordination. MiMo, on the other hand, has caught up quickly through rapid iteration and efficiency gains. In terms of benchmark scores, the two are now effectively on equal footing. The competition is no longer just about raw model performance. The focus is shifting toward real-world productivity and the ability to complete complex tasks end to end. In that context, Xiaomi’s latest model signals something bigger. China is no longer just catching up. It is now building its own presence in the top tier of global AI. by Ju-baek Shinㅣ[email protected]