In real-world tests under equivalent model and quantization conditions, the token generation speed of the MLX solution is typically 10%–30% faster than the Metal backend of llama.cpp.