On September 22, Xiaomi officially launched its new generation large model MiMo-V2.6, releasing two native multimodal versions, Pro and Flash, at once. Just a few days earlier, Xiaomi had live-streamed the RL post-training process of V2.6. Now, the moment of announcement has arrived—Pro version scored 46 points on the Artificial Analysis Comprehensive Intelligence Index (AA Intelligence Index), leaving Kimi K3 and Qwen3.8Max behind, and taking the position of the current strongest open-source model. However, Xiaomi was very honest, stating that it still lags behind top closed-source models like Claude Fable5.1 and GPT-6Astra.

image.png

What really makes competitors sweat is the price. The pricing of V2.6 is locked in with its predecessor V2.5: Flash cache hit costs only 0.02 yuan, and Pro is just 0.025 yuan. The off-peak price of Flash is the same as V4.1 Flash, and the unit prices for input and output are even cheaper. The V4Pro that was mentioned in comparison has been completely beaten by the market, and beating it is no longer a significant achievement. As a result, it's clear that MiMo-V2.6 has become the new benchmark for AI cost-effectiveness—only Astra, Fable, and Meta's Muse1.3Max and GPT-5.6Sol Max can outperform it, and even those models need to run at maximum reasoning chain intensity to catch up. Regarding the cost-effectiveness of Xiaomi, the official provided a set of data: about 60 times that of overseas AI.

Along with the new model, the MiMo Desktop desktop client and membership subscription plan were also launched. Subscribers can access both Pro and Flash of V2.6, and they can also configure their own API Key to connect to the client. Even more impressive, MiMo Desktop provides an UltraSpeed high-speed mode for the Pro version, increasing the inference speed up to 20 times, specifically designed for scenarios requiring strong real-time interaction and particularly sensitive to response latency.