Chen Danyan's StartLux Emerges as Strong Contender in China's Large AI Model Race

A new player has entered the competitive field of large AI models in China, challenging established giants with a smaller yet highly capable offering. StartLux, formerly known as Yuandian Xinghui and led by Chen Danyan, a veteran entrepreneur and early internet pioneer, introduced its StartLux-V1.0-27B-Preview model, which has only 27 billion parameters. Despite its smaller size, it secured second place in the China Academy of Information and Communications Technology (CAICT) MCP specialized benchmark test, narrowly trailing DeepSeek-V4-Pro, a 1.6 trillion parameter model, by just 1.3 percentage points. The CAICT MCP benchmark evaluates AI models based on their ability to perform real-world tasks across six categories including navigation, web search, browser automation, financial analysis, code management, and 3D design. Unlike traditional benchmarks focused on response quality, MCP emphasizes practical task completion and multi-tool collaboration, reflecting the growing importance of AI agents capable of complex interactions. StartLux’s model demonstrated strong performance, ranking first in location navigation, financial analysis, and browser automation. In a financial analysis comparison, StartLux accurately identified trading days and closing prices, outperforming a competing model by providing verifiable and traceable results. In a browser-based flight search task, StartLux completed the process in less than half the time of its competitor and found a significantly cheaper ticket, illustrating its operational efficiency. The success of StartLux’s smaller model is attributed to advanced post-training techniques rather than sheer parameter size. Building on the Qwen3.6-27B model, StartLux applies automated task-focused training that enhances the model’s ability to understand tasks, select tools, execute multi-step processes, and verify results. This approach, called Auto Research, enables the model to autonomously adapt strategies based on real-world feedback, marking a departure from the traditional scaling law that prioritizes larger models. Chen Danyan, known for co-founding Shanda Network and pioneering shared software concepts in China, has returned from retirement to focus on local AI models. He argues that local models will disrupt cloud-based AI markets by offering more cost-effective, personalized, and privacy-conscious solutions that run directly on consumer-grade hardware. This vision aligns with recent industry trends where major companies like Meta, Google, and NVIDIA are also exploring local AI models, though StartLux is among the first to commercialize this approach in China. StartLux’s leadership team combines entrepreneurial experience with technical expertise. Co-founder and CTO Guo Quanwei, a computer science Ph.D. with a background in privacy-preserving machine learning and AI for science and finance, leads the technical roadmap. Another co-founder, Luo Yongxiang, brings financial and marketing expertise from his tenure at Morgan Stanley Asia. Looking ahead, StartLux plans to release its first-generation local intelligent solutions for enterprises and individual users within the year, aiming to simplify deployment and usage to the level of installing common software. This approach could broaden access to AI capabilities without requiring specialized hardware or technical knowledge. The emergence of StartLux underscores a broader shift in the AI landscape, where efficiency, adaptability, and local deployment are gaining prominence alongside traditional large-scale cloud models. As China’s AI ecosystem evolves, StartLux represents a significant step toward diversified AI solutions that balance performance with practicality.