LLM quantization

Four bits, then a cliff

How to choose LLM quantization for production: why 4-bit is the default, where 1-bit collapses, and why your own eval set is the real deployment gate.