Primary use stated or inferred cautiously from official documentation.
N
NVIDIA / nvidia/Nemotron-3.5-Parse
Nemotron 3.5 Parse
NVIDIA model released on October 17, 2026 to turn page images into structured text, semantic classes, boxes and reading order. It combines C-RADIOv4-SO400M with a ten-layer mBART decoder and 630,800,496 unique parameters. It accepts an RGB image and a task prompt; PDFs must be converted into page images. It does not generate images or replace a general-purpose assistant.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | nvidia/Nemotron-3.5-Parse |
|---|---|
| Model type | OCR and documents · Multimodal |
| Access model | Open weights |
| License | OpenMDW 1.1 |
| Deployment | Transformers · vLLM with Nemotron plugin |
| Release | 17 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Image · Text |
| Exit | Text |
| Context window | 16,384 positions shared between image and text |
| maximum output | 9,000 configured tokens; evaluation with 8,744 |
| Reasoning | Visual extraction, not a conversational assistant |
| Published tools | OCR |
| Structured Outings | Boxes, classes and reading order; postprocessing available |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Open weights |
| Verified platforms | Hugging Face · Transformers · vLLM |
Document, table and chart extraction and data preparation for search and RAG. The reference recommends images up to 1,664 × 2,048 pixels. Coordinates are content-normalized and must be scaled to the evaluated image. The helpers preserve reading order and merged cells when exporting HTML; a flat Markdown table does not retain all spans.
The configuration declares 16,384 positions shared by the visual prefix and text, not 16,384 available conversation tokens. generation_config sets 9,000 new tokens; examples and OCR evaluations use 8,744. These are not two different universal limits. The reference code requires trust_remote_code: review it and pin a revision before execution. Check omissions, figures, privacy and reading errors with your own documents.
03 / CAPABILITIES
What it can do
Orientation
Document, table and chart extraction and data preparation for search and RAG. The reference recommends images up to 1,664 × 2,048 pixels. Coordinates are content-normalized and must be scaled to the evaluated image. The helpers preserve reading order and merged cells when exporting HTML; a flat Markdown table does not retain all spans.
Context
16,384 positions shared between image and text · 9,000 configured tokens; evaluation with 8,744
Tools and integration
OCR
Access
Hugging Face · downloadable checkpoint
Input modalities
ImageText
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Downloadable weights under OpenMDW 1.1, with commercial and non-commercial use subject to its terms. The tokenizer has MIT attribution and dependencies retain their licenses. We have not verified a commercial API or token pricing: numeric costs remain empty. Self-hosted inference requires a budget for hardware, memory, operations and security; availability for download does not imply free inference.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Results published by NVIDIA, not reproduced by Inferama. ParseBench covers 2,078 pages: the highlighted score uses triple pass with region union, not plain triple pass. IndicVisionBench OCR uses 876 pages and ten languages. MOSCAR uses 8,111 synthetic pages and 163 language/script entries, corresponding to 162 language codes. Its character bag F1 ignores order and does not guarantee accuracy on real scans, handwriting or degraded images.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| ParseBench Overall (triple-pass union) | 75,20 | Higher is better | View source ↗ |
| ParseBench Tables (triple-pass union) | 84,67 | Higher is better | View source ↗ |
| ParseBench Charts (triple-pass union) | 64,40 | Higher is better | View source ↗ |
| ParseBench Content Faithfulness (triple-pass union) | 87,96 | Higher is better | View source ↗ |
| ParseBench Semantic Formatting (triple-pass union) | 68,28 | Higher is better | View source ↗ |
| ParseBench Visual Grounding (triple-pass union) | 70,67 | Higher is better | View source ↗ |
| IndicVisionBench OCR ANLS-char | 83,73 | Higher is better | View source ↗ |
| IndicVisionBench OCR CER | 16,00 | Lower is better | View source ↗ |
| MOSCAR Character Bag F1 | 99,51 | Higher is better | View source ↗ |
| MOSCAR CER | 1,35 | Lower is better | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Nemotron 3.5 Parse
Release of weights and OCR references; the repository's creation in September is not used as its release date.
07 / SOURCES
Traceability
MODEL CARDNemotron 3.5 Parse · ficha y lanzamientohuggingface.co↗
CONFIGURATIONContexto visual y textualhuggingface.co↗
CONFIGURATIONGeneración predeterminadahuggingface.co↗
ASSESSMENTParseBench · métricas y revisioneshuggingface.co↗
ASSESSMENTOCR multilingüe · metodologíahuggingface.co↗
ASSESSMENTOCR multilingüe · resultadoshuggingface.co↗
LICENSEOpenMDW License Agreement 1.1openmdw.ai↗
BENCHMARKParseBench Overall (triple-pass union)huggingface.co↗
BENCHMARKIndicVisionBench OCR ANLS-charhuggingface.co↗
BENCHMARKMOSCAR Character Bag F1huggingface.co↗