NVIDIA adds document-image search to Nemotron 3.5 retrieval models

Search reaches the information inside document images

NVIDIA announced its Nemotron 3.5 retrieval collection on October 21, adding models that help AI applications find and prioritize evidence in text and document images. The release combines three Embed variants, an 8B Rerank model and a document parser. It targets developers building search and assistants over collections containing charts, tables and scanned pages.

The announcement packages several downloads with different release dates. The Embed and Rerank cards list October 20; the Parse card lists October 17. These are available model components for developers to integrate, with commercial and noncommercial use identified in their cards.

Source 1Source 2Source 3Source 4Source 5Source 6

Finding a page and deciding its relevance are separate jobs

Embed converts each document into a numerical representation that a search system can store and compare with a user's question. A document can be text, a page image or both. Queries remain text-only. The 8B model accepts up to 32,768 tokens, so the input budget still constrains how much material an application can supply.

Rerank operates after that initial search. It evaluates a question together with each candidate document and returns a relevance score, allowing the application to reorder the shortlist. Its output is a score rather than a written answer. A higher score indicates a closer match to the question, not proof that the document's claim is correct.

In a hypothetical engineering assistant, the first stage could locate specification pages containing a relevant table; the second could prioritize the page addressing the requested component. Our reading is that this gives developers more control over which evidence reaches the answering model, while leaving the answer and its verification to the wider application.

Source 2Source 3

A migration path for existing text search

NVIDIA says the 8B BF16 model retains the language component of Nemotron 3 Embed 8B and produces identical text embeddings. The 1B BF16 card likewise describes a shared text embedding space with its corresponding predecessor. That is a useful distinction for teams that have already indexed a large text collection: the visual upgrade does not inherently require replacing those compatible text representations.

The compatibility claim is specific. The 1B card documents 2,048-dimensional vectors and supports shorter vectors only with normalization after truncation. Our inference is that an existing deployment still needs to preserve its matching representation and preprocessing settings. A shared family name alone does not make every stored vector interchangeable.

Source 2Source 4

The compact option has a hardware boundary

The 1B NVFP4 variant uses lower-precision weights and activations in both its language and vision components. Its implementation guidance specifies vLLM and hardware with NVFP4 support, including NVIDIA Blackwell GPUs. Developers using Transformers or Sentence Transformers are directed to the BF16 model instead.

The BF16 1B card lists Linux and NVIDIA Ampere, Hopper and Blackwell compatibility. It also documents separate software environments: its Transformers examples use version 5.20.0, while the vLLM examples use Transformers 5.19.0 with vLLM 0.32.0. The card warns that Transformers 5.17.0 through 5.19.0 can produce CUDA errors or incorrect image outputs with Flash Attention 2. These dependencies make runtime selection part of the deployment decision.

Source 4Source 5

Parse supplies structure when pixels must become data

Nemotron 3.5 Parse addresses another part of the workflow. Its approximately 631-million-parameter model turns page images into text with element classes, bounding boxes and reading order. The card describes uses including scanned PDFs, slides, forms and reports. Postprocessing can convert supported table or chart text into formats such as HTML, JSON or CSV.

That output serves a different purpose from a search vector: an application can work with extracted content and its location on the page. Accuracy remains task-dependent. NVIDIA explicitly cautions that its multilingual OCR results on simply rendered synthetic pages do not guarantee accuracy on complex scans, handwriting or degraded images. Broad language evaluation is therefore not evidence that every real document will parse reliably.

Source 6

Token savings are a vendor result, not a complete operating bill

NVIDIA reports that adding its reranker reduced agent inference-token use by about 14% with the 8B embedding model and 13% with the 1B model in its experiments. Its evaluation used ViDoRe v3, BRIGHT and BrowseComp-Plus, holding the agent, tools, tasks and execution policy fixed while changing retrieval configurations. AI Toolcast has not reproduced these results.

Our analysis is that fewer agent tokens do not establish an equivalent reduction in total spending. Retrieval and reranking themselves require computation, and an application must store and serve its index. The model cards identify commercial use under OpenMDW 1.1, but deploying downloadable weights still leaves those infrastructure costs with the operator. The practical opportunity is more selective evidence retrieval, with economics dependent on the actual workload.

Source 1Source 3Source 5