The 4B-pretrained model can not recognize video input.

Thanks for your great work!

I download the 4B-pretrained and 8B-pretrained checkpoints, and convert them to huggingface format.
Then I feed a video into them, respectively. They are told to describe the video.
However, the 4B model always outputs "The image is a photograph...", while the 8B model correctly outputs "The video appears to be a still shot from a video..."

So, does the 4B model after pretraining can not recognize video input?