Qwen stumbles on SVG code

My take: this is less a Qwen-specific failure than a reminder that SVG is a nasty edge case for LLMs. It asks for valid code, spatial planning, and renderable output at the same time, and most models still break when those constraints stack up.

• –SVG generation is not just syntax completion; it needs path logic, coordinate accuracy, and layout consistency.
• –Recent SVG-focused benchmarks and papers show performance drops fast as scenes get more compositional or visually exacting.
• –Qwen's coding and reasoning strengths do not automatically transfer to vector graphics unless the training data includes lots of SVG-like examples.
• –A render-validate-refine loop or specialized SVG toolchain is usually more reliable than asking a model to freehand the whole asset.
• –For developers, this is a good reminder to separate “can write code” from “can produce faithful visual code.”