NVIDIA opens Nemotron 3.5 Parse for document OCR, charts and structured layouts

NVIDIA has published Nemotron 3.5 Parse, a specialist open-weight document image model with 630,800,496 trainable parameters. Its primary task is extracting structured information from rendered pages, including text, semantic layout classes, reading order and spatial bounding boxes. It is not a new version of Nemotron 3.5 Super VL or Bolt, and it should not be evaluated as a general-purpose chat assistant.

The architecture combines a C-RADIOv4-SO400M image encoder, an adapter and a ten-layer mBART decoder. Dynamic rectangular image preprocessing preserves document structure, while task-specific tokens identify regions, charts and text. Output bounding boxes use coordinates normalized against the page dimensions. Tables and chart data need downstream conversion or postprocessing rather than being guaranteed universally machine-readable without validation.

The model card reports ParseBench mean 75.20 and a charts subscore of 64.40. It also includes IndicVisionBench and MOSCAR evaluations; the latter uses relatively simple synthetic text renderings and character-overlap measures. These figures are NVIDIA-reported rather than independent Open Model Weights results. They do not establish field accuracy for scanned invoices, handwriting, rotated pages or highly complex charts.

NVIDIA distributes the weights under OpenMDW 1.1 and documents separate MIT-governed tokenizer attribution. Transformers and a special vLLM integration are described, including optional one-token speculative decoding. Serving with custom repository code requires dependency review. Teams should verify exact artifact revisions, licensing, document-level quality and safe handling of uploaded pages before production use. This release offers a promising small parser for RAG ingestion without eliminating normal data-extraction safeguards.