Quantization is a common way to make deep learning models more efficient for deployment on devices with limited memory and compute. Post-training quantization is especially attractive because it can be applied to an already-trained model without needing the original data or a costly retraining pass. A new paper on arXiv introduces VLCQ, a post-training quantization approach that incorporates variable-length coding.

The abstract describes VLCQ as a method that uses variable-length coding during quantization, which could allow different weights to be represented with different numbers of bits. This is in contrast to fixed-length quantization schemes that assign the same bit width to every weight. The paper's abstract is truncated in the source, so specific experimental results or comparisons to other methods are not available from the provided text.

Overall, the contribution appears to be a new way to compress neural networks after training, with the potential to reduce model size while avoiding the need for retraining. Further details on performance and trade-offs would require reading the full paper.