Xiaomi MiMo-V2-Flash开源：能力比肩标杆闭源模型Claude 4.5 Sonnet

Group 1 - Xiaomi officially announced the open-source release of Xiaomi MiMo-V2-Flash, a MoE model with a total parameter count of 309 billion (15 billion activated), achieving top 2 in global open-source model benchmarks [1] - The model features innovations such as Hybrid attention architecture and multi-layer MTP inference acceleration, resulting in a code capability comparable to the closed-source model Claude 4.5 Sonnet, but at only 2.5% of its inference cost and with a 2x increase in generation speed [1] - Xiaomi MiMo-V2-Flash outperformed DeepSeek V3.2 and K2-Thinking in most evaluation benchmarks, reducing parameter count by 50% to 67%, and achieving low cost and high speed, with preliminary capabilities to simulate the world [1] Group 2 - The next generation of intelligent agent systems is envisioned not merely as "language simulators" but as true "intelligent agents" that understand and coexist with the human world [2] - There is a shift in agent execution capabilities from merely "answering questions" to "completing tasks," incorporating memory, reasoning, autonomous planning, decision-making, and execution abilities [2] - Unified multimodal perception is essential for understanding the physical world, which will enhance integration with smart devices like glasses [2]