Qwen 27B's memory needs reportedly reduced from about 60GB to 6GB

A post says a Prism ML team reduced Alibaba's Qwen 27B's memory needs from about 60GB to 6GB by storing its numbers with 1.5 bits instead of 16, while keeping 95% of its performance. It says the model runs on a Raspberry Pi at 6GB. The post also quotes Emad Mostaque saying AI models have “pretty much reached the efficiency of a human brain” and likening the Pi's energy use to the brain's.