Multiverse Computing Shrinks AI Model While Boosting Performance

Multiverse Computing researchers have developed a compression method called Quantization-Aware Healing, aimed at reducing the cost of deploying large artificial intelligence models without the usual loss in capability. The approach combines model pruning and low-bit quantization with supervision from the original model, challenging the assumption that smaller systems must inevitably perform worse than their full-sized counterparts.

The team reduced a 120-billion-parameter model to 60 billion parameters and compressed it to 4-bit precision. The smaller model outperformed the original full model in seven of nine benchmark tests, according to the reported results. The findings suggest that using the source model to guide post-compression recovery can cut memory and computing requirements while preserving, and in some tasks improving, model performance.