Qwen3.6-35B-A3B outperforms Claude Opus 4.7 on image generation tasks in informal testing

Alibaba's Qwen3.6-35B-A3B, released alongside Claude Opus 4.7, was subjected to Willison's established—if tongue-in-cheek—visual generation test. Running a 20.9GB quantized version locally via LM Studio on an M5 MacBook Pro, the model produced an SVG drawing of a pelican on a bicycle that correctly rendered the bicycle frame and included contextual details like clouds and a caption. Anthropic's new flagship model stumbled on the same task, generating a distorted bicycle frame in its initial attempt and producing a similar error even when invoked with maximum reasoning mode.

Concerned that the labs might deliberately optimize for his quirky benchmark—a worry he has addressed before—Willison deployed an undisclosed secondary test asking for an SVG of a flamingo on a unicycle. Again, Qwen3.6-35B-A3B delivered output with greater visual personality and technical finesse, including a well-placed SVG comment. Opus 4.7 produced a competent but unremarkable illustration lacking decorative elements.

Willison framed the findings within the proper context of his benchmark's limitations. He emphasized that the pelican test originated as satire about the absurdity of model comparison itself, yet acknowledged that historically, pelican quality has tracked with general model utility improvements over time. However, he explicitly rejected the inference that a quantized 35B model now outperforms Anthropic's latest proprietary system in practical capability, noting skepticism that the correlation holds in this case.