4 ms·

My guess they are trying lower quantization. I have not noticed I recently begun trying qwen 3.5 locally.