Comment by jantse

Thanks for sharing! Did you observe a speed difference between ollamas mlx version and the mlx-community/Qwen3.8-27B-4bit from HF ran with mlx_vlm.generate (with MTP)? Or is it the same?