I feel like Google is using the bottom-most tier of their models or at least a heavily quantized version for generating video summaries. I've been testing gemma 3 for data extraction and it's essentially useless.