Nemotron 3.5 Super VL parses PDFs, tables, charts, and videos without requiring a separate vision model, achieving an 80.8 percent score on the CharXiv chart benchmark