According to @_avichawla, 8 LLM precisions trade accuracy for memory, enabling FP8, INT8, and NF4 gains and KV cache cuts for longer contexts.