Xiaomi's MiMo V2.6 Pro has become the highest-ranked open-weight AI model on the Artificial Analysis composite index, scoring 46 points, according to a Chinese media roundup of October model releases.
What is drawing as much attention as the score is the training budget. Reports said the model's reinforcement learning stage took six days and cost roughly 3.47 million dollars, a fraction of what overseas frontier labs spend, with estimates putting comparable overseas training runs at 20 to 60 times the cost.
The result reinforces a theme of China's October model wave: rapid iteration at very low cost. Bloomberg Intelligence noted that the benchmark gap between the best Chinese and American models has narrowed to about 3 percent, down from 15 percent at the start of the year and 9 percent in May.
OpenRouter usage data cited in the same roundup showed domestic Chinese models accounting for 57 to 67 percent of platform tokens by mid-September, up from 6 to 13 percent in February, a reversal that took barely a quarter.
Analysts caution that Chinese models still trail on long-horizon agents, high-difficulty reasoning, top-end chips and ecosystem tooling, with one estimate putting the remaining gap at about eight months. Even so, the combination of near-frontier performance, open weights and radical cost efficiency is pulling global developers toward Chinese models for both experimentation and production workloads.
Source: Chinese tech media roundup citing Artificial Analysis, OpenRouter and Bloomberg Intelligence, October 2026.
