Granite 4.1 3B SVG Pelican Gallery

2026-05-04 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

IBM recently released its Granite 4.1 family of large language models (LLMs) under an Apache 2.0 license, available in 3B, 8B, and 30B parameter sizes. Concurrently, Unsloth published a collection of 21 GGUF-encoded, quantized variants of the 3B Granite 4.1 model, with file sizes ranging from 1.2GB to 6.34GB, totaling 51.3GB. An experiment was conducted using these Unsloth GGUF files to prompt "Generate an SVG of a pelican riding a bicycle" across different quantized sizes. The results showed no discernible pattern linking output quality to model size; all generated SVGs were of poor quality, with even the smallest model producing a better bicycle image than larger variants.

Key takeaway

For AI Engineers evaluating quantized LLMs for creative content generation, you should not assume that larger quantized models will inherently produce higher quality outputs. Your focus should be on empirical testing across various quantization levels and model sizes for your specific use case, as smaller variants might surprisingly outperform larger ones in certain tasks, challenging conventional wisdom about model scaling.

Key insights

Quantization level does not guarantee improved output quality for specific creative tasks.

Method

The experiment involved prompting 21 different GGUF-quantized variants of the Granite 4.1 3B model with the same SVG generation request to compare output quality against model size.

In practice

• Test multiple quantization levels for creative tasks.
• Evaluate model performance beyond just size metrics.

Topics

• Granite 4.1 LLMs
• IBM Foundation Models
• GGUF Quantization
• Unsloth Collection
• SVG Generation

Best for: Machine Learning Engineer, Prompt Engineer, AI Engineer

Related on AIssential

• Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7 — Local, quantized models can surprisingly outperform larger proprietary models in specific creative tasks.
• LLMs Generate Kitsch — LLMs systematically generate kitsch due to their training, despite often outperforming human-generated works in ratings.
• Lessons from GGUF Evaluations: Ternary Qwen3.5, Bricked Minimax — Model robustness to quantization varies significantly, necessitating empirical evaluation beyond perplexity metrics.
• GLM-5.1: Towards Long-Horizon Tasks — GLM-5.1 autonomously generates complex animated SVG/CSS, demonstrating advanced multimodal reasoning and self-correction capabilities.
• LLM Watch Weekly: When Scale Isn't Enough — Scaling alone cannot overcome data biases or fundamental reasoning gaps in LLMs and VLMs.
• How Large Language Models Are Actually Built, Everything I Learned From Stanford’s CS229 LLM… — Data, evaluation, and systems are the true differentiators in LLM quality, not just architecture.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.