How Wenxin X1.1 Tops China’s LLMs on the New SuperCLUE-CPIF Benchmark

SuperCLUE-CPIF Benchmark Launch

Recently, the Chinese Precise Instruction Following benchmark (SuperCLUE-CPIF) was officially released. Wenxin X1.1 scored 75.51 points, ranking first among domestic large models and leading both in task type and instruction count categories.

The evaluation covered ten models, including GPT-5(high), DeepSeek-V3.2-Exp-Thinking, Claude-Sonnet-4.5-Reasoning, Gemini-2.5-Pro and others. SuperCLUE-CPIF focuses on assessing large language models’ ability to precisely follow complex, multi-constraint Chinese instructions and to transform natural-language commands into outputs that satisfy all requirements.

Results show Wenxin X1.1 at the top with 75.51 points, followed by DeepSeek-V3.2-Exp-Thinking (73.98) and Hunyuan-T1-20250822 (65.82) among domestic models.

Wenxin X1.1 is built on the Wenxin 4.5 model and uses an iterative mixed-reinforcement-learning framework. This approach simultaneously improves general-task performance and agent-task performance, while iterative self-distillation data generation continuously enhances overall model quality.

At the WAVE SUMMIT 2025 demo, X1.1 demonstrated strong capabilities in complex writing tasks, combining internal knowledge, web-search tools, and deep reasoning to produce fact-accurate, well-structured, logical, and elegant content. In multi-step service scenarios such as handling diverse user issues on a bike-sharing platform, the model followed business processes, invoked tools autonomously, and adapted to user emotions.

As one of the earliest Chinese companies to invest in large-model R&D, Baidu leverages a full-stack self-developed ecosystem of chips, frameworks, models, and applications. Thanks to PaddlePaddle-Wenxin joint optimization, X1.1 improves factuality by 34.8%, instruction compliance by 12.5%, and agent performance by 9.6% compared with the original X1.