The Voodoo dynamic quant method goes MIT: gradient descent picks the per-tensor quant layout
The author of Voodoo Quant, who kept the method private for two months after claiming SOTA at aggressive quant levels on small Qwen3.5 GGUFs, released the toolset under MIT at github.com/curvedinf/voodoo-dyn-quant. The method is simpler than the community guessed: run every quant level for every tensor at once, freeze the candidate quant weights from llama.cpp's ggml conversion, and train a single scalar gate per tensor per quant level, with softmax keeping gradient flowing to all levels and an annealed tau forcing each tensor to settle on one choice. For anyone shipping GGUFs this is a reproducible recipe rather than another mystery quant, and the author explicitly handed it over because he does not have time to scale it.
↳ Follow the thread