ADVANCING LLM COMPRESSION WITH ADDITIVE QUANTIZATION

Summary from the TLDR AI newsletter

This project suggests a new compression technique for LLMs using Multi-Codebook Quantization (MCQ). The method, adapted from Additive Quantization, sets a new standard in LLM compression, achieving higher accuracy at lower bit counts.