Vista-LLaMA is an advanced video language model aimed at improving video understanding. It minimizes the generation of text unrelated to video content by maintaining equal distance between visual and language tokens, regardless of the length of the generated text. This method omits relative positional encoding when calculating the attention weights between the computational vision and text tokens, making the influence of visual tokens more prominent during text generation. Vista-LLaMA also introduces an ordered visual projector that projects the current video frame onto the tokens in the language space, capturing temporal relationships within the video while reducing the reliance on visual tokens. The model has demonstrated significantly superior performance compared to other methods on multiple open-source video question-answering benchmark datasets.