Domestic Models Surge: Doubao Joins Global Elite, Xiaomi MiMo Top in Math Reasoning

Chinese large language models are quickly evolving from mere followers to active collaborators — and even pioneers in certain fields.

On March 30, the Chinese large model benchmark assessment SuperCLUE published its March 2026 results. The evaluation covered 22 major models from China and abroad, testing six core capabilities including mathematical reasoning, scientific reasoning, and code generation. The findings indicate that domestic models, led by "Doubao," have now secured spots in the global top tier.

Global View: Overseas Closed-Source Models Still Lead, Doubao Close Behind

In the overall rankings, overseas closed-source models continued to show formidable technical prowess:

Top Three: Anthropic's Claude-Opus-4.6, Google's Gemini-3.1-Pro, and OpenAI's GPT-5.4 took first, second, and third place globally.

Domestic Highlight: ByteDance's Doubao (Doubao-Seed-2.0-pro) claimed the top domestic position with 71.53 points. It not only stays in the global top tier but has also cut the gap with GPT-5.4 to a mere 0.95 points.

Intelligent Agent Breakthrough: In the dimension of agent task planning, Doubao outperformed several overseas models, breaking into the global top five.

Xiaomi's Performance: MiMo-V2 Series Excels in Mathematical Reasoning

As a major smartphone maker venturing into AI, Xiaomi Group's MiMo series delivered consistent performance in this evaluation:

Mathematics Star: MiMo-V2-Pro placed among leading closed-source models with 60.67 points, scoring an outstanding 84.03 in mathematical reasoning tasks.

Double Model Listing: Alongside the Pro version, the open-source MiMo-V2-Flash also appeared on the list, demonstrating strong potential in areas such as code generation.

Open Source Track: Domestic Models Claim Comprehensive Leadership

Unlike the tight race in closed-source, domestic models have established a clear dominance in the open-source arena:

Top Three: Domestic open-source models including Kimi-K2.5-Thinking and Qwen3.5-397B secured the top three spots on the open-source leaderboard.

Downward Strike: Evaluation data reveals that domestic open-source models have clearly outperformed their overseas rivals, making them a go-to choice for developers worldwide.

Conclusion: Moving from Parameter Wars to Real-World Performance

As the March 2026 ranking shows, Chinese large models have moved beyond just understanding Chinese language contexts. They are now going head-to-head with global leaders in demanding areas like logical reasoning and code generation. With Doubao rising in the ranks and Xiaomi MiMo