Native MCP:

For Qwen 35B enabling native MCP on MLX models slows it down by 10%.

For Qwen 27B enabling native MCP on MLX models speeds token generation up almost exactly 1.5x.

(all tested on M5 pro).