The core behaviors of inference serving software occur at runtime, including real model loading into GPU memory, concurrent request batch scheduling, dynamic KV cache management, and throughput and latency fluctuating with load.