Startlux-27b challenges DeepSeek in AI benchmark, marks new era for local models

A recent report by the China Academy of Information and Communication Technology's Institute of Artificial Intelligence revealed that StartLux-V1.0-27B-Preview, developed by Shanghai YuanDian Starlight Science and Technology Co., Ltd., outperformed DeepSeek-V4-Flash in the Trusted AI Large Model Benchmark Test (MCP Special Test). With 27B parameters, the model ranked second, achieving a comprehensive score of 39.25, surpassing DeepSeek-V4-Flash-0731 and Step-3.7-Flash. It also outperformed Qwen-3.6-27B by 5.34 percentage points. The breakthrough was attributed to StartLux's post-training directional enhancement and the use of AI Training AI (Auto Research) for optimization. This marks the first time a Chinese local Agent model has employed this method. The test evaluated six tasks, including location navigation, web search, and financial analysis, emphasizing multi-tool collaboration and real-world interaction. Industry experts note that the focus is shifting from parameter scale to practical problem-solving, data security, and cost control. Global tech giants like Google, Meta, and NVIDIA are accelerating their local model development, signaling a growing market for such models. StartLux plans to launch its first local intelligent solution this year, with ongoing research into new architectures like diffusion language models.