Qwen3.6-27B-Q6_K handles whimsical SVG prompts locally

Hot take: this is a useful local multimodal/synthetic-image benchmark because it stresses both instruction following and structured output, not just pretty one-off generations.

• –The prompt set is intentionally diverse, covering animals, clothing, action scenes, and a complex multi-stage composition, which gives a better read on consistency than a single showcase image.
• –The reported speeds are fairly tight, so the runtime looks stable rather than cherry-picked.
• –Because the post does not include side-by-side outputs or failure cases, it is hard to judge whether the model is actually drawing clean SVGs or merely producing passable concept sketches.
• –Hardware details are thin, so the performance numbers are informative but not directly comparable across other local setups.
• –For the LocalLLaMA crowd, the interesting part is that a 27B Q6_K quant seems capable enough for playful image-generation tasks, which broadens the practical use case beyond text and code.