“A recent OpenAI model with 120 billion parameters uses 4-bit quantization for most of its layers, allowing it to fit within 60 gigabytes of memory.”