Nemotron 3.5 Parse
nvidia/Nemotron-3.5-Parse
Open Source · chat · open-weights
Open
Alert me on changes
Context
—
Max output
—
Weights
Open
API $/1M
—
Modalities
text · image
Released
03 Sept 2026
License: openmdw-1.1 · nvidia/Nemotron-3.5-Parse
AI summary
● machine-written
Nemotron 3.5 Parse: NVIDIA's multimodal document understanding model
Nemotron 3.5 Parse is an open-weights multimodal model developed by NVIDIA for document understanding, information retrieval, data extraction, and multimodal data-curation workflows. It uses dynamic rectangular NaViT preprocessing, grouped-query self-attention, and includes chart-aware output capabilities. The model is intended for parsing and extracting structured information from documents and images.
What's new
- Dynamic rectangular NaViT preprocessing for document image handling
- Grouped-query self-attention with 16 query heads and 8 key/value heads
- 72,505-token multilingual tokenizer
- Chart-aware output through dedicated <class_Chart> token
- Released under OpenMDW-1.1 license
Best for
Document understanding and parsingInformation retrieval from multimodal documentsData extraction from structured documentsMultimodal data curation workflows