StartLux's local model StartLux-V1.0-27B-Preview ranks second in CAICT MCP benchmark

A 27B-parameter local model from StartLux, the Shanghai lab founded by Chen Danian, scored 39.25 on the CAICT Institute of Artificial Intelligence's MCP benchmark, placing second. StartLux-V1.0-27B-Preview finished ahead of DeepSeek-V4-Flash-0731, a 284B model, and roughly one percentage point behind the 1.6-trillion-parameter DeepSeek-V4-Pro.

Built on Qwen3.6-27B, the model's gains come entirely from post-training, the team says. It calls the method Auto Research: about 70% of experiment execution was handled by AI, with the stated goal of recursive self-improvement (RSI) running on a personal computer.

Original sources (Chinese)