“Amjad Masad highlights a recent development with DeepSeek-OCR where processing a screenshot of text is more token-economical for a model's context window than processing the raw text itself.”