Qwen3.5 9B through Ollama (GGUF and MLX), llama-bench, mlx-lm and rapid-mlx on an M3 Ultra. Measured tokens per second, memory, and what MTP changes.