I had the same reaction with Deepseek V4 ! It would be more useful as a vision model