Chinese large models are accelerating their transformation from "followers" to "co-players" and even "leaders".

On March 30, the Chinese large model benchmark evaluation SuperCLUE released the latest results for March 2026. This evaluation included 22 mainstream models from both domestic and international sources, covering six core tasks such as mathematical reasoning, scientific reasoning, and code generation. The results show that domestic models represented by "Doubao" have successfully entered the global top tier.

image.png

Global Perspective: Overseas Closed-Source Models Still Dominate, Doubao Follows Closely

In the overall ranking of this evaluation, overseas closed-source models still demonstrated strong technical strength:

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

Domestic Highlight: ByteDance's Doubao (Doubao-Seed-2.0-pro) secured the top spot domestically with 71.53 points. It not only remains in the global top tier but has also narrowed the gap with GPT-5.4 to just 0.95 points.

Intelligent Agent Breakthrough: In the intelligent agent task planning dimension, Doubao even surpassed some overseas models and entered the global top five.

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

As a representative of major smartphone manufacturers entering AI, Xiaomi Group's MiMo series performed steadily in this evaluation:

Mathematics Star: MiMo-V2-Pro ranked among the top closed-source models with 60.67 points, achieving an impressive score of 84.03 in mathematical reasoning tasks.

Double Model Listing: In addition to the Pro version, the open-source version MiMo-V2-Flash also made the list, showing remarkable potential in specific scenarios like code generation.

Open Source Track: Domestic Models Achieve "All-Round" Leadership

Compared to the fierce competition in the closed-source field, domestic models have shown a dominant advantage in the open-source track:

Top Three: Domestic open-source models such as Kimi-K2.5-Thinking and Qwen3.5-397B took the top three positions on the open-source ranking list.

Downward Strike: Evaluation data shows that domestic open-source models have significantly outperformed their overseas counterparts, becoming a new favorite among global developers.

Conclusion: From "Parameter Competition" to "Practical Capabilities"

From this ranking released in March 2026, it can be seen that Chinese large models no longer merely focus on understanding the Chinese language context. Instead, they are directly competing with global top models in hard-core areas like logical reasoning and code generation. With Doubao moving up in position and Xiaomi MiMo