Chinese AI model DeepSeek has gone viral over the weekend, causing stock markets to crash, AI labs to issue statements, and even US President Donald Trump to say that its success was a “wake up call” for the US AI industry. But while the world is waking up to the capabilities of a Chinese AI labs, DeepSeek isn’t the only Chinese AI company of note.
China’s AI landscape has been booming, with leading companies delivering state-of-the-art models that compete on the global stage. This article highlights the top 10 AI language models from Chinese labs, based on their performance, architecture, and innovations. Data is based on the latest benchmarks and releases.
Top Chinese AI models
1. DeepSeek-R1 by DeepSeek AI
- Architecture: Mixture of Experts (MoE)
- Parameters: 685 billion
- Tokens: 14.8 trillion
- Performance:
- MMLU: 90.8%
- MMLU-Pro: 84.0%
- GPQA: 71.5%
- Release Date: January 2025
- Highlights: This model leads in Chinese AI innovation with unparalleled general knowledge and professional accuracy. Its strong MMLU score of 90.8% places it as the best-performing model on this list.
2. Hunyuan-Large by Tencent
- Architecture: Mixture of Experts (MoE)
- Parameters: 389 billion
- Tokens: 7.0 trillion
- Performance:
- MMLU: 89.9%
- MMLU-Pro: 60.2%
- GPQA: 42.4%
- Release Date: November 2024
- Highlights: Tencent’s Hunyuan-Large excels in general tasks but lags behind in domain-specific performance compared to other models. It is optimized for efficient token utilization.
3. Doubao-1.5-Pro by ByteDance
- Architecture: Mixture of Experts (MoE)
- Parameters: 300 billion
- Tokens: 9.0 trillion
- Performance:
- MMLU: 88.6%
- MMLU-Pro: 80.1%
- GPQA: 65.0%
- Release Date: January 2025
- Highlights: Known for its balance of compactness and accuracy, Doubao-1.5-Pro shows high performance on professional tasks, second only to DeepSeek-R1.
4. MiniMax-Text-01 by MiniMax
- Architecture: Mixture of Experts (MoE)
- Parameters: 456 billion
- Tokens: 7.2 trillion
- Performance:
- MMLU: 88.5%
- MMLU-Pro: 75.7%
- GPQA: 54.4%
- Release Date: January 2025
- Highlights: MiniMax-Text-01 shines in language understanding with a mid-range parameter count, making it ideal for resource-constrained scenarios.
5. Qwen2.5-Max by Alibaba
- Architecture: Mixture of Experts (MoE)
- Parameters: 325 billion
- Tokens: 20.0 trillion
- Performance:
- MMLU: 87.9%
- MMLU-Pro: 69.0%
- GPQA: 60.1%
- Release Date: January 2025
- Highlights: With one of the largest token datasets, Qwen2.5-Max is well-suited for long-context understanding and large-scale generative tasks.
6. Kimi k1.5 by Moonshot AI
- Architecture: Dense
- Parameters: 500 billion
- Tokens: 15.0 trillion
- Performance:
- MMLU: 87.4%
- GPQA: 51.5% (MMLU-Pro unavailable)
- Release Date: January 2025
- Highlights: As one of the few dense models on this list, Kimi k1.5 delivers solid generalist performance, emphasizing dense parameter utilization over modular architectures.
7. Yi-XLarge by 01-AI
- Architecture: Mixture of Experts (MoE)
- Parameters: 2 trillion
- Tokens: 20.0 trillion
- Performance:
- MMLU: 85.1%
- GPQA: 48.2% (MMLU-Pro unavailable)
- Release Date: May 2024
- Highlights: With the largest parameter count, Yi-XLarge focuses on scalability but has slightly lower efficiency and performance compared to compact models.
8. SenseNova 5.0 by SenseTime
- Architecture: Mixture of Experts (MoE)
- Parameters: 600 billion
- Tokens: 10.0 trillion
- Performance:
- MMLU: 84.8%
- GPQA: 42.9% (MMLU-Pro unavailable)
- Release Date: April 2024
- Highlights: Positioned as a reliable model for mid-tier tasks, SenseNova 5.0 balances performance and efficiency.
9. GLM-4 by Zhipu AI
- Architecture: Dense
- Parameters: 130 billion
- Tokens: 10.0 trillion
- Performance:
- MMLU: 83.3%
- GPQA: 39.9% (MMLU-Pro unavailable)
- Release Date: January 2025
- Highlights: With a focus on lightweight and dense architecture, GLM-4 is designed for low-resource environments and targeted tasks.
10. Step-2 by StepFun
- Architecture: Mixture of Experts (MoE)
- Parameters: 1 trillion
- Tokens: 13.0 trillion
- Performance:
- MMLU: 82.9%
- MMLU-Pro: 63.0% (GPQA unavailable)
- Release Date: July 2024
- Highlights: Step-2 boasts high parameter efficiency and token diversity, suitable for a broad spectrum of applications.
Conclusion
China’s AI labs are pushing the boundaries of what language models can achieve, with models like DeepSeek-R1 and Doubao-1.5-Pro leading in professional performance. The variety in architecture and specialization ensures a robust ecosystem of tools tailored to diverse needs, from large-scale professional tasks to resource-constrained applications.