PPL vs model size - safe to assume larger size == better accuracy regardless of UD vs non-UD?
If we could have a table or diagram to better understand model size vs accuracy that would be very helpful, thank you.
Am I right to assume that Qwen3-32B-UD-Q5_K_XL.gguf is superior to Qwen3-32B-Q5_K_M.gguf? Both are 23.2GB. How big of a difference in accuracy is there?
UD is generally preferred over Standard Quant from Unsloth.
Based on the metrics I have seen, comparing Main Branch, Unsloth UD, and Bartowski's; they all have varying degrees of good and bad. If you solely look at Delta Log Probability: Main is the worst, Unsloth UD is better, and Bartowski is slightly more narrow still. However, that is just one metric. Unsloth leads in inference speed and memory footprint, making it ideal for resource-constrained environments.
The key takeaway is: No single method dominates across all metrics; however, to answer your question-- In general, UD should be preferred over the standard quant as the metrics show increased performance per Gib.