Quantized Granite 4.1 3B fails the pelican-on-bicycle SVG test across all 21 variants

IBM shipped its Granite 4.1 LLM family under Apache 2.0 in 3B, 8B, and 30B sizes, with training methodology documented by team member Yousaf Shah. Unsloth followed up with 21 GGUF-quantized builds of the 3B model spanning 1.2GB to 6.34GB, totaling 51.3GB across the collection.

Simon Willison used the full quantization spread to run a long-planned experiment: prompting each variant to generate an SVG of a pelican riding a bicycle, looking for a quality-versus-size signal. The result was a flat line of failure. No quantization tier produced meaningfully better output than another, and all of them were poor. The takeaway is narrow but useful: for this particular generative task, quantization granularity is not the variable that matters, and Granite 4.1 3B is not the right base model to probe it with.