Breaking the 1.58-bit Barrier for Ternary LLMs
Researchers have developed BITCOS, a new compression technique for ternary large language models that breaks the theoretical 1.585-bit barrier by exploiting the actual distribution of weights in these models, which contain up to 51.5% zeros. The method achieves 1.485 bits per weight on sparse models and delivers up to 1.28× faster matrix-vector multiplication and 1.27× faster inference throughput on GPUs compared to standard packing approaches. BITCOS includes optimized unpacking implementations for modern CPUs and GPUs, making it practical for deployment across multiple hardware platforms.