A study by ByteDance reveals that a 7B model can effectively answer questions about long, image-heavy documents, outperforming larger models. This approach, which focuses on question-answering rather than transcription, may reshape training methods for language models.