StartLux 27B Model Rivals Trillion-Parameter Flagships in CAICT Test

According to a report by Wall Street CN, the China Academy of Information and Communications Technology (CAICT) recently conducted a specialized test based on the Model Context Protocol (MCP), which evaluates AI models' ability to integrate with external tools and data. Shanghai-based StartLux's local model StartLux-V1.0-27B-Preview achieved the second-best overall performance.

Despite having only 27 billion parameters, the model outperformed DeepSeek-V4-Flash and came within about 1% of the 1.6 trillion-parameter DeepSeek-V4-Pro in overall score. In Agent tasks, including financial analysis and browser automation, StartLux's model ranked first, showcasing its strength in real-world applications.

The company announced plans to launch commercial solutions based on this model within the year. This highlights how efficient, compact models can rival much larger flagships, offering cost-effective AI for businesses.