October 9, 2026 – SuperCLUE has released a Chinese-language benchmark for agent-based terminal programming, with Xiaomi’s MiMo V2.6 Pro taking the top overall score.
The evaluation focuses on how well large models perform when driving an agent framework to handle Chinese-language requirements. It covers requirement understanding, task planning, tool use, file editing, error diagnosis and repair, and final delivery. All models were tested on the same Claude Code agent framework to keep the comparison consistent.
Among the remaining participants, Kimi ranked second, GLM third, and Hy4 fourth. Xiaomi’s MiMo V2.6 Flash came in fifth.
MiMo V2.6 Pro scored 64.65 points at an average cost of 1.52 yuan per task, the highest score in the test. MiMo V2.6 Flash scored 53.54 points at 0.67 yuan per task, the lowest cost among the models evaluated. By comparison, DeepSeek-V4.1-Flash (max) scored 50.51 points at 1.86 yuan per task, while Hy4-Preview (high) scored 55.56 points at 11.61 yuan per task. MiMo V2.6 Pro outperformed Hy4-Preview on score while costing far less per task, showing a strong performance-to-cost balance on this task set.
Across the full task flow, DeepSeek-V4.1-Flash stood out for time efficiency, while MiMo V2.6 Pro delivered a better overall trade-off between high scores and execution time.
Many developers are already using coding agents as daily assistants rather than expecting them to write complete programs from scratch. They are more commonly used for generating scripts, diagnosing errors, and fixing bugs. In Chinese-language scenarios, some overseas models still struggle with instruction understanding, giving domestic models a clear adaptation advantage.
Agent-based terminal programming is still evolving, and no model can fully replace professional programmers yet. But the latest results show that domestic large models are making rapid progress on hard engineering tasks and are increasingly competitive with leading international products.
