The total model is 3B, right? And you have one text expert and one image expert. In every model forward, the input will be split two set, the text tokens are fed into text expert and image tokens are fed into image expert. So, it seems that all param are calculated in model forward. If I miss anything, please correct me. Thanks