Xiaomi Is Suddenly Competing With the World’s Top AI Labs

Xiaomi's MiMo-V2.6-Pro has recently taken the top spot in Artificial Analysis' open-weight model rankings, joining ranks with some of the most powerful AI models available. The model scored a 46 on the Intelligence Index, matching Grok 4.7 High. This puts it ahead of Kimi K3 Max (44), GLM-5.3 Max (45), and GLM 5.3 Flash (42). The result is remarkable, as MiMo 2.6 Pro and Grok 4.7 arrived around the same time in late September, yet MiMo's open-weight model achieves the same score while remaining accessible.

Comparing the two models reveals distinct strengths. MiMo 2.6 Pro outperforms Grok 4.7 High in Terminal-Bench, SciCode, Humanity's Last Exam, CritPt, and long-context reasoning tasks. Conversely, Grok leads in AA-Briefcase, GDPval, AutomationBench, and GDP.pdf. Therefore, while they share an overall score, their individual capabilities vary.

Cost is another area where MiMo shines. Benchmarking shows that MiMo 2.6 Pro costs $0.13 per task for an Intelligence Index task, while Grok 4.7 High costs $2.73. This makes MiMo roughly 21 times more affordable, with an estimated full Intelligence Index task costing $207 versus $3,881 for Grok.

However, Xiaomi's MiMo 2.6 Pro does not excel in raw generation speed. Grok 4.7 High outperforms in output speed, with faster token generation and shorter times to the first token and answer. Xiaomi does offer a faster version, MiMo-V2.6-Pro-UltraSpeed, tailored for latency-sensitive tasks.

In comparison to Kimi K3 Max, MiMo 2.6 Pro again leads in overall index score. While Kimi performs better in certain long-context tasks, MiMo's overall score and performance on coding, science, and reasoning tests give it the edge.

GLM-5.3 Max is also notable, with a score of 45. MiMo 2.6 Pro surpasses GLM in several areas, including science, complex reasoning, and long-context tasks. The models differ notably in cost per task; MiMo costs $0.13, while GLM charges $2.01.

The model architecture behind MiMo 2.6 Pro is unique. It utilizes a Mixture-of-Experts framework with 1 trillion total parameters, of which 42 billion are active during inference. Supporting a 1-million-token context window, it can generate up to 128,000 output tokens. Released under the MIT license, MiMo is commercial-use friendly and allows modification.

Designed for complex tasks, long-running projects, research, cybersecurity, and agent workflows, it supports tool calling, web search, structured output, streaming, and context caching. Xiaomi claims that reinforcement learning has improved MiMo 2.6 Pro's software engineering performance, increasing its DeepSWE v1.1 score from 58.4 to 72.6 after extensive training involving approximately 750,000 trajectories and costing around $2.62 million.