Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 21 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-20
NVIDIA Nemotron 3.5 Parse is a dedicated, downloadable roughly 631-million-parameter document image parser for OCR, reading order, semantic layout regions and bounding-box extraction. It is not the larger Nemotron 3.5 Bolt or Super VL conversational model. The first-party repository records an exact 40-character SHA, one recognized main Safetensors weight artifact and NVIDIA OpenMDW v1.1 licensing, checked by Open Model Weights on 20 October 2026. The source config declares 16,384 positional tokens, but this is not a validated capacity for arbitrary PDF documents. NVIDIA's published ParseBench scores are producer evaluations; OMW has not run this model or independently measured OCR accuracy, latency or VRAM.
WHY IT MATTERS
RAG ingestion usually fails before generation if document text, regions or reading order are extracted incorrectly. A purpose-built parser can support more inspectable downstream source chunks.
A 631M-class checkpoint creates a different hardware design space than enormous vision-language systems, but its actual deployment footprint must still be measured with the image encoder and page preprocessing.
MODEL-SELECTION VIEW
Nemotron 3.5 Parse focuses on recognizing document text and classifying visual page regions such as lines, charts or tables while retaining normalized bounding boxes and reading order. These are intermediate extraction artifacts intended for search indexing, content conversion or quality review, not direct answers to arbitrary user questions.
Applications must convert geometry into page pixels and decide how to combine regions, preserve headings and assign references. A coordinate rectangle does not itself prove that the text inside it is correctly recognized, and a page-level summary may hide errors that become important in downstream RAG.
NVIDIA describes C-RADIOv4-SO400M for image encoding connected through a visual-sequence adapter to a ten-layer mBART decoder. Its instructions support dynamically shaped rectangular pages and task tokens that separate text extraction from layout regions. These details are operational contracts rather than interchangeable components of any generic vision chat pipeline.
Source-based inference examples cover Transformers and a specialized vLLM integration. Pin the documented code revision and processor behavior, including image resizing, language tokenization and coordinate normalization. The config's 16,384-token limit is a source field and not an independently tested promise for all document dimensions.
A page with columns, footnotes or overlapping graphical labels can be more difficult than a clean paragraph. Test region sequences and coordinate overlap directly, then reconstruct the desired reading order before indexing. Parsing failure can create apparently coherent but factually incomplete chunks if a chart label is separated from its legend.
For tables, check merged cells, column headers, numerical units and carry-over across pages. For charts, compare extracted labels and data points against the original image where postprocessing is required. Treat the parser as one component in an extraction pipeline rather than automatic proof of machine-readable truth.
NVIDIA's card reports ParseBench mean 75.20 and a chart-category score of 64.40. It also presents OCR evaluations such as IndicVisionBench and MOSCAR; the latter involves simpler synthetic text renderings. Each evaluation has its own input normalization, datasets and scoring method.
These are developer-produced results, not independently replicated Open Model Weights measurements. Do not compare the headline score with a different benchmark or claim uniform accuracy across scripts, handwriting, low-resolution scans or complex tables. Sample held-out documents from the deployment domain and compare against manually reviewed transcriptions.
The official repository includes a recognized Safetensors model artifact and an OpenMDW Agreement version 1.1 license file. NVIDIA additionally distinguishes MIT attribution for the tokenizer. Preserve these specific obligations; neither OpenMDW nor MIT for a component implies all associated inference libraries or model weights use the same terms.
The publisher describes roughly 631 million trainable parameters. Weight-only arithmetic does not include processor tensors, intermediate image features, activations or serving allocations, and can substantially underestimate memory under multi-page or concurrent workloads. Benchmark exact resolution, precision and batch sizes before promising local-device fit.
Use document-level source IDs and page coordinates when passing OCR output into a downstream embedding or reranking system. Keep original images for audit and retain extraction confidence or human corrections where available. The derived text must not be treated as a trusted instruction channel for an agent.
Our assessment: choose Parse when geometry-aware extraction is the bottleneck, the 8B Embed checkpoint when initial semantic document discovery is the bottleneck, and the distinct 8B Rerank checkpoint when the relevant evidence appears in the retrieved set but is ordered poorly. These three model roles are complementary, not substitutes.
IMPLEMENTATION & EVALUATION
Practical source-grounded evaluation plan for the exact nvidia/Nemotron-3.5-Parse revision, not an Open Model Weights runtime or benchmark execution.
Check the official model SHA, processor, task tokens, OpenMDW and tokenizer attribution separately.
Archive the reproducible revision manifest and one representative page input with its recognized regions.
An unreviewed custom-code update can change output schemas or introduce parsing regressions.
Choose multi-column pages, chart labels, forms, tables, low-resolution scans and multilingual cases.
Store expected reading order, region geometry, original page image and human-reviewed text.
A plausible OCR transcript can omit an entire page region without an obvious syntax error.
Measure text errors, missed bounding boxes, header association and table cell structure per document type.
Track manually checked errors and stop deployment if important numeric or coordinate fields fail.
Publisher benchmark aggregates cannot predict the accuracy of an unfamiliar report format.
Carry page IDs and original image references through chunking, embedding and access-control checks.
Demonstrate correct citations, private-document access rules, extraction fallbacks and revision rollback.
A visually convincing generated answer can conceal an OCR error introduced many steps earlier.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
No. It is a specialist document image parser for OCR text, structural regions and coordinates; answering arbitrary user questions requires other components.
NVIDIA documents chart and table-related outputs, but extracting reliable numeric structures can require postprocessing and manual checks on complex pages.
NVIDIA's model card identifies 630,800,496 trainable parameters, commonly described as 631 million. That count is not a measured runtime memory requirement.
The checked config contains 16,384 positional tokens; this should not be advertised as a tested capacity for any arbitrary number of PDF pages or image pixels.
NVIDIA describes commercial and non-commercial uses subject to OpenMDW v1.1. Review the full weight agreement plus tokenizer and custom-runtime dependencies.
No. ParseBench, MOSCAR and language-specific results currently cited on this page are publisher-reported; local-document error rates require a separate evaluation.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Nemotron 3.5 Parse 631M as a 0.6308B total (publisher-declared)-parameter document images → ocr text, layout and bounding boxes open-weight model in the Nemotron 3.5 Parse family, with NVIDIA listed as the developer. The checked source repository provides 1 recognized weight artifacts, format evidence for Safetensors and a verified context window of 16,384 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares NVIDIA OpenMDW License Agreement v1.1. Commercial and non-commercial use declared by NVIDIA; OpenMDW 1.1 checkpoint-specific conditions govern. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: described-in-model-card.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial and non-commercial use declared by NVIDIA; OpenMDW 1.1 checkpoint-specific conditions govern
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.