An MSR blog introduces our efforts on low-bit LLMs