On February 6, 2026, Apple Inc. released its cutting-edge AI research, revealing that the Qwen3-Coder model—despite its relatively compact parameter size—has outperformed GPT-5 in generating application interfaces, thanks to an innovative fine-tuning method. Spearheaded by Apple’s UICoder team, the study underscores that conventional fine-tuning techniques fall short in the specialized field of UI design. To overcome this challenge, Apple engaged 21 expert designers who contributed detailed feedback through written critiques, hand-drawn sketches, and code adjustments. Tests showed that the model trained with “sketch-based feedback” achieved the best results, outperforming GPT-5 after fine-tuning with just 181 sketch annotations. The research also discovered that concrete visual improvements tend to foster greater agreement among designers compared to abstract ratings.
