Release of Voodoo Dynamic Quant Toolset
A new open-source methodology for model quantization, titled “Voodoo Dynamic Quant” (VQ), has been released under the MIT license. The toolset, hosted on GitHub by developer curvedinf, aims to provide a method for creating “dynamic quants”—a process where different quantization levels are applied to individual tensors within a model to optimize performance and file size.
According to the developer, the methodology differs from existing “static” quantization techniques. While static methods apply fixed quantization levels to specific types of tensors, VQ utilizes gradient descent to optimize the per-tensor quantization layout. The process involves running all available quantization levels simultaneously and using a single scalar gate per tensor to determine which levels minimize loss for a specific target file size.
The developer reports that the method has shown exceptional results on smaller model sizes, particularly at aggressive, low-bit quantization levels. In comparative testing against Unsloth Dynamic 3.0 (UD 3.0), the developer noted that while UD 3.0 performs better at mid-to-high quantization levels, VQ demonstrates superior performance at more aggressive levels.
The developer characterized the project as “research grade” and noted that it has not yet been studied at larger model scales. While the toolset is currently configured for the Qwen model architecture, the documentation suggests it can be adapted for other architectures.
Technical Implementation and Limitations
The implementation of VQ relies on a training epoch where candidate quantization weights are frozen, and only the scalar gates are trained. The optimization uses a loss function based on the Kullback–Leibler divergence (KLD) between the mixed-quant logits and a reference BF16 checkpoint.
Technical discussions on Reddit highlighted potential challenges regarding the use of gradient descent on non-continuous quantization levels. The developer addressed this by explaining that the process uses a “tau level” that is annealed to help the training freeze into singular predominant selections.
Furthermore, the developer acknowledged a discrepancy in numerical precision. Currently, VQ utilizes Torch numerics, which are more precise than the underlying llama.cpp library used by competitors like Unsloth. While this results in better performance within the Torch environment, the developer noted that the advantage may not translate fully to llama.cpp until the numerical differences are addressed.
Developments in Autonomous Agent Harnesses
In a separate development within the open-source AI space, a developer has released “Artificium,” a general agent harness designed for long-term autonomous work and self-improvement. The project, hosted on GitHub, is intended to manage its own context and working memory, allowing it to perform tasks such as coding or acting as a personal assistant.
To test the harness, the developer conducted a 63-hour experiment involving the Riemann hypothesis. According to the developer’s report, the experiment utilized a Qwen3.8-27B model with 4-bit quantization running on a single RTX 3090. The run processed over 50 million input/output tokens.
The developer stated that the objective of the experiment was not to solve the mathematical problem, but to test the ability of a 27B model to maintain focus and avoid hallucinations during long-running, autonomous tasks. The experiment’s metrics and logs have been made available on Hugging Face for public inspection.