Nemotron 3.5 Parse: Benchmarks, Deployment and Document Parsing Tradeoffs

Nvidia released Nemotron 3.5 Parse on October 17, a downloadable document model that extracts text, reading order, layout labels and bounding boxes from page images. For developers building document search or retrieval-augmented generation, its most useful detail is the choice between faster parsing and additional passes over tables and charts.

The model has approximately 631 million parameters, including its prediction head. That puts the release in a different deployment category from a large general-purpose vision model, although parameter count alone does not establish its memory requirements or production cost.

What Nemotron 3.5 Parse produces

The model card describes image input and structured text output. Bounding boxes locate extracted content on the page; layout labels distinguish the kinds of content being read. The weights use OpenMDW-1.1, which permits commercial and noncommercial use, with separate notices applying to included components.

A document pipeline needs these relationships. If a parser detaches a table heading from its values, the retrieval system can find the right page and still hand the answering model misleading context. Keeping location and structure gives a downstream application more information to preserve, inspect and cite.

Benchmarks: choose the inference profile

Nvidia reports the following ParseBench results and throughput on 2,078 prepared page images, using one H100 per model. These are vendor measurements, not independent production tests.

Nemotron 3.5 Parse profile	ParseBench score	Pages per second
Single pass	72.74	15.42
Double pass	74.92	11.65
Triple pass (region union)	75.20	10.62

The comparison methodology makes an important distinction. Other models' accuracy figures come from published results, while their speeds were measured locally. Protocols and scorer versions can differ. That makes a narrow cross-model score lead less conclusive than a fully matched accuracy experiment. The refinement merge also retains differing full-page and crop alternatives, which are included in the reported scores.

Timing includes model calls and dependent crop work, but excludes initial page preparation, startup, warmup, final rendering and file writes. The table therefore describes a serving workload rather than the time required to turn an arbitrary PDF into a finished document.

Within Nvidia's own three profiles, double pass captures most of the reported accuracy improvement. Triple pass adds 0.28 points while reducing throughput by about 9%. That is a useful starting point for evaluation, rather than a reason to make the most expensive mode the default.

Why additional passes help

The pipeline documentation explains the mechanism. Single pass reads the full page. Double pass revisits predicted table and chart regions as crops. Triple pass adds a separate detection stage; the region-union option combines regions found by the detector and full-page parser.

This spends more computation on content that benefits from a closer view. The released pipeline also converts predictions into document formats and applies structural rules. Those rules are heuristics and can make mistakes on unfamiliar documents.

For an invoice extraction system, the practical question is whether the extra pass fixes the fields that matter. A slightly better aggregate score is less useful if totals, column associations or footnotes still need manual repair. Evaluate accepted documents and correction effort alongside pages per second.

Deployment and API access

Nvidia provides Transformers and vLLM paths. Its serving guide pins a Python 3.12 environment and specifies BF16 on an H100 as the reference configuration. The vLLM route uses the repository's serving extras and model plugin, with one-token speculative prediction enabled automatically.

The documented server exposes an OpenAI-compatible interface. That describes a self-hosted endpoint; the release does not establish a hosted API price. Infrastructure, preprocessing and review remain part of the cost.

Keep coordinate and class tokens in raw predictions for postprocessing, and use the model-specific instructions rather than assuming a generic vision-model quick start is sufficient. The serving guide recommends reducing concurrency on GPUs with less memory, but does not make the reference throughput a promise for those GPUs.

The useful deployment decision is how much rereading each workload deserves. A parser that exposes that tradeoff clearly gives teams a better basis for testing whether additional compute buys fewer document errors.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn model releases and engineering updates into clear stories that developers and founders can understand and share.

Book a 30-minute call.