Because llama.cpp is primarily optimized for single-user or low-concurrency scenarios, Ollama generally underperforms vLLM in high-concurrency throughput.