A New Community SVG Generation Prompt for Benchmarking Local LLMs
A Reddit user has proposed a new casual benchmark prompt for testing the spatial reasoning and code generation capabilities of LLMs by asking them to generate an SVG of a horse on a blue bicycle in the desert with a camel in the background. The prompt was used to compare models like Qwen3.8-27b, Sol 5.6, and Qwen3.6-35B. While informal, such community-driven prompts serve as practical, hands-on heuristics to quickly evaluate how well different LLMs handle complex spatial layouts and precise vector graphic generation. They highlight the gap between frontier models and quantized local models in executing multi-object instructions. The prompt intentionally includes a typo ("bycicle") to test the model's robustness, and the comparison involves different quantization formats like Unsloth Dynamic (UD-Q4_K_XL). Generating valid, visually coherent SVG code requires both strong coding skills and a deep understanding of spatial relationships.
## BACKGROUND
SVG (Scalable Vector Graphics) generation is a common test for LLMs because it requires the model to translate text descriptions into precise coordinate-based code. Evaluating local LLMs often involves testing quantized versions, such as Unsloth Dynamic (UD) quantizations, which optimize model size by assigning varying precision levels to different neural network layers. Frontier models like OpenAI's Sol 5.6 represent the high-end baseline for these reasoning tasks.