Voodoo Quant optimizes Qwen3.5 0.8B and 2B with lower KLD than Unsloth

A new mixed-precision optimization technique called Voodoo Quant has been applied to Qwen3.5 0.8B and 2B models, claiming a 95% reduction in Kullback-Leibler divergence (KLD) compared to Unsloth Dynamic 2.0.

• Voodoo Quant optimizes every tensor individually rather than blocks of tensors, using a different methodology for precision selection.
• The technique demonstrates competitive performance across both Torch and Llama.cpp graph structures, whereas Unsloth's method shows significantly poorer performance in Torch.
• This suggests Unsloth may overfit for Llama.cpp, while Voodoo offers a more generalized optimization approach.
• The author notes that "2 bit" quantization appears to be the optimal setting for this technique.

The creator aims to expand this research to larger models like Qwen3.6 27B and DeepSeek v4 flash to improve their utility on lighter hardware.