It should be lower, but I haven't optimized it too hard on disk size. It's about iteration latency for LLM's.