IBM Granite 4.1 LLM Family and Unsloth 3B GGUF Variants
Benchmark the 21 Granite 4.1 3B GGUF variants against your SVG generation workload to assess performance and output quality.
Run a benchmark of the 21 GGUF models on your SVG pipeline, record latency and output fidelity, and decide whether to adopt the 3B variant for production.
Summary
IBM has just released the Granite 4.1 family of large language models, available in 3B, 8B and 30B parameter sizes under an Apache 2.0 license. The release includes a detailed training description from team member Yousaf Shah and a suite of 21 GGUF‑encoded, quantized variants of the 3B model, each ranging from 1.2 GB to 6.34 GB in size.
Unsloth’s collection of 21 GGUF files totals 51.3 GB. The author tested these variants by prompting each model to generate an SVG of a pelican riding a bicycle. The results were uniformly poor, with no clear correlation between model size and output quality; the smallest model produced the best bicycle, while the largest produced a vague pelican‑like shape.
The experiment highlights that, at least for this specific SVG task, larger GGUF variants do not guarantee better visual fidelity. The release provides a useful benchmark for developers interested in lightweight, quantized LLMs for creative generation tasks.
Key changes
- Granite 4.1 LLM family released in 3B, 8B and 30B sizes
- Apache 2.0 license granted for all models
- Unsloth released 21 GGUF‑quantized variants of the 3B model
- Variant sizes range from 1.2 GB to 6.34 GB
- Total size of all variants is 51.3 GB
- SVG generation test shows no quality scaling with size
- Smallest variant produced best bicycle representation
- Largest variant produced a vague pelican‑like shape