Voodoo Quant claims much lower error for small Qwen3.5 models
Voodoo Quant has released new GGUF versions of the Qwen3.5 0.8B and 2B models for lighter local use. It uses , giving more accurate number formats to the parts of a model judged more important.
Unsloth Dynamic 2.0 makes this choice for blocks of tensors, while Voodoo Quant optimizes every tensor separately with a new method. Its creator reports a 95% KLD improvement over Unsloth Dynamic 2.0 in one comparison.
However, that striking figure depends on the measurement setup, and the published graphs also show smaller, more realistic differences. Torch and produce separate results because they arrange their calculations differently, with Torch generally considered the more precise of the two.
Key points
- New GGUF sets are available for Qwen3.5 0.8B and 2B.
- More precise number formats are kept for important model parts.
- Voodoo Quant tunes every tensor instead of groups of tensors.
- The creator reports a 95% KLD improvement in one comparison.
- Torch and results must be compared separately because their calculation structures differ.