A dark horse has emerged in China's AI race as StartLux-V1.0-27B-Preview secured second place in the CAICT MCP specialized benchmark test, trailing DeepSeek-V4-Pro by only 1.3 percentage points. The model, developed by StartLux (formerly Yuandian Xinghui), achieved a comprehensive score of 39.25% despite having 27 billion parameters compared to DeepSeek-V4-Pro's 160 billion. StartLux also outperformed Qwen 3.6 by 5.34 percentage points at the same parameter scale and ranked first in location navigation, financial analysis, and browser automation.

Founder and CEO Chen Danian, known as the "godfather" of Chinese programmers and co-founder of Shanda Networks, returned from a decade of retirement to bet on local models. At the Shanda Innovation Institute's 18th anniversary, he declared eight non-consensus views for the AI era, predicting local models will destroy the cloud market, catch up to Claude within three years, and capture 80% market share.

StartLux focuses on commercializing small local models that run directly on consumer-grade PCs without cloud dependency. In a financial analysis test against Claude Sonnet 4.6, StartLux calculated a Microsoft stock investment return of $47,499.09 (90.00%) versus Claude's $47,254 (89.02%), demonstrating how minor accuracy differences translate to significant financial impact. The MCP-Universe benchmark evaluates real-world task completion across six scenarios including code repository management and 3D design.