The story isn't only that Voodoo Dynamic Quant is now MIT licensed. It's that choosing the quantization level for each tensor, long approached as a static analysis problem, is being tackled with a backward pass and gradient descent. The project was announced two months ago by its author, who kept the methodology private; now the GitHub repository exposes the tools for building dynamic quants independently.

The mechanism is more elegant than much of the community had speculated. Voodoo Quant runs all quantization levels for every tensor in parallel, freezes the candidate weights provided by the library underlying llama.cpp, and trains a single scalar gate per tensor per quantization level. As training progresses, a tau parameter is annealed to push the system toward predominant selections, while softmax continues to distribute gradient to less-chosen levels. The loss measures the KL divergence between the mixed-quant logits and a BF16 reference checkpoint, rewarding both similarity and closeness to the target file size.

The result is research-grade: it has not been studied on larger models. The data cited by the author concern smaller Qwen3.5 models in GGUF format, where VQ stands out especially at the most aggressive quantization levels. For people running local inference on low-VRAM machines, this is the most sensitive point: aggressive quants let larger models fit in memory, but usually impose a sharp quality loss. Picking the level for each individual tensor can reduce that loss without changing the hardware.

The comparison with Unsloth Dynamic 3.0 helps frame the paradigm shift. In the author's tests, UD3 remains better at medium-to-high levels, while VQ wins at more aggressive ones. According to the author, UD3 is an evolution of the static analysis techniques that currently dominate; Voodoo Quant is the first method, to his knowledge, to use a backward pass to optimize the per-tensor quantization layout. The computational cost is higher: more powerful hardware is needed than for static analysis. But the principle is potentially more efficient at maximizing performance, because it replaces many repeated benchmarking iterations with a differentiable comparison.

Structurally, opening the code moves dynamic quantization from a proprietary black box to a reproducible component of the local pipeline. In on-premise or air-gapped environments, an MIT tool makes it possible to verify the process, adapt it to different architectures, and keep control over calibration data and checkpoints. Anyone evaluating on-premise deployment knows the question is not only which LLM to use, but how to prepare it for VRAM and TCO constraints; AI-RADAR publishes analytical frameworks at /llm-onpremise to compare such trade-offs. The ball now passes to those with GPU capacity and the willingness to dig into the code.