StartLux's 27B parameter model, StartLux-V1.0-27B-Preview, ranked 2nd overall in the MCP test conducted by the China Academy of Information and Communications Technology (CAICT).
This test evaluates Agent capabilities across six types of tasks based on practical scenarios. The StartLux model recorded an overall score of 39.25%.
This score is only 1.3 points behind DeepSeek-V4-Pro, which has 1.6 trillion parameters. Among models of a similar scale with 27B parameters, it outperformed Qwen 3.6 by 5.34 points.
StartLux was founded by Chen Danyan, who previously served as a co-founder of Shanda Network. The company specializes in the commercialization of local LLMs that run directly on consumer PCs without relying on the cloud.
In individual tasks, the model took 1st place in three categories: location navigation, financial analysis, and browser operation.
Source: A dark horse enters China's AI race: StartLux (Hacker News Frontpage, 2026-09-03)