NVIDIA Releases Nemotron 3.5 Embed, Rerank and Parse Open Retrieval Models

NVIDIA has expanded its open AI model portfolio with five Nemotron 3.5 retrieval models designed to help AI agents find, rank and extract information from enterprise documents. Announced on October 21, 2026, the release covers three essential tasks in AI-powered information retrieval: embedding, reranking and document parsing. The models are intended to help businesses build retrieval-augmented generation (RAG) systems that can search internal knowledge bases and provide large language models with relevant evidence before generating answers.

The new collection includes Nemotron 3.5 Embed in 8-billion- and 1-billion-parameter versions, a compressed 1-billion-parameter NVFP4 variant, Nemotron 3.5 Rerank with 8 billion parameters, and Nemotron 3.5 Parse with 631 million parameters. NVIDIA says the models are openly available for commercial use, giving developers more flexibility to build and deploy retrieval systems on infrastructure they control. The company also reports that adding its reranker reduced agentic token costs by approximately 13%–14% in selected tests while improving retrieval accuracy.

Key Takeaways

• Five retrieval models released: NVIDIA introduced three embedding variants, one reranker and one document parser.
• Designed for AI agents: The models help agents retrieve relevant information from enterprise knowledge bases and documents.
• Multimodal retrieval: The embedding and reranking models can work with text and visual content, including tables, charts and diagrams.
• Open commercial availability: Developers can access the models through Hugging Face and NVIDIA’s model-serving ecosystem.
• Lower reported inference costs: NVIDIA says reranking reduced agentic token costs by roughly 13%–14% in its evaluations.
• Document parsing: Nemotron 3.5 Parse is designed to extract information from PDFs and other complex documents.
• Greater deployment flexibility: Organizations can customize retrieval workflows and run models on supported infrastructure.

What Are NVIDIA’s Nemotron 3.5 Retrieval Models?

AI systems often need to search large collections of documents before answering a question. A business assistant, for example, may need to locate relevant information in product manuals, financial reports, contracts, technical diagrams and internal policies.

A general-purpose language model does not automatically know which documents contain the answer. A retrieval system must first find useful evidence and then supply it to the model.

NVIDIA’s Nemotron 3.5 Retriever collection addresses this process through three complementary capabilities.

Model	Size or variant	Primary purpose
Nemotron 3.5 Embed	8B BF16	High-quality multimodal retrieval
Nemotron 3.5 Embed	1B BF16	Smaller-footprint multimodal search
Nemotron 3.5 Embed	1B NVFP4	Compressed embedding for high-throughput workloads
Nemotron 3.5 Rerank	8B BF16	Reorder search results by relevance
Nemotron 3.5 Parse	631 million parameters	Extract structured information from documents

Source: NVIDIA’s model announcement and accompanying technical information.

These models perform different jobs rather than competing to provide the same capability. Embedding helps find candidate information, reranking improves the order of those candidates, and parsing turns complex documents into information that downstream systems can process.

How Nemotron 3.5 Embed Improves AI Search

Turning documents into searchable representations

Embedding models convert text and other supported content into numerical representations that can be compared during a search. Instead of relying only on exact keyword matches, a retrieval system can use these representations to find information that is semantically related to a user’s question.

For example, someone asking an enterprise assistant about a company’s employee travel reimbursement policy might receive relevant passages even if the documents use different wording from the question.

NVIDIA’s new embedding models extend this capability to visually rich documents. Relevant information may appear not only in paragraphs but also in tables, charts, infographics, product datasheets and technical schematics.

Three variants for different deployment needs

The 8-billion-parameter BF16 model is designed for quality-focused multimodal retrieval. The 1-billion-parameter BF16 version offers a smaller model for organizations that need multimodal search with a more compact deployment footprint.

The 1-billion-parameter NVFP4 version uses a compressed numerical format intended to support large-scale indexing and high-throughput retrieval on compatible NVIDIA hardware.

The choice between these variants involves trade-offs. Larger models may provide stronger retrieval performance in some workloads, while smaller or compressed models can reduce memory requirements and improve serving efficiency. Actual performance depends on the hardware, corpus, indexing strategy and evaluation task.

Nemotron 3.5 Rerank: Finding the Most Relevant Results

An initial search can return many potentially useful documents without identifying which ones best answer the question. A reranker evaluates the relationship between the query and each candidate result, then reorders the results to place the most relevant evidence higher.

NVIDIA’s Nemotron 3.5 Rerank is an 8-billion-parameter model designed to score query-document pairs, including candidates containing text and visual information.

This step can be especially useful for AI agents that repeatedly search a large knowledge base. If the most relevant evidence appears near the top of the results, an agent may need fewer additional searches and less processing before it can complete a task.

NVIDIA reports lower token costs

NVIDIA evaluated its retrieval pipeline on benchmarks including ViDoRe v3, BRIGHT and BrowseComp-Plus. The tests examined how agents search a corpus, review retrieved evidence and refine subsequent queries.

According to NVIDIA, adding Nemotron 3.5 Rerank reduced agentic token costs per query by approximately 14% when paired with the 8B embedding model and 13% when paired with the 1B version. The company also reported improved retrieval accuracy in both configurations.

In a comparison using the 8B embedding model, NVIDIA said its reranker achieved approximately 0.9 more nDCG@10 points than Qwen3-VL-Reranker-8B at a comparable agentic token cost.

These figures are vendor-reported benchmark results. They do not establish that every application will achieve equivalent savings, because performance depends on the workload, model configuration and retrieval infrastructure.

Nemotron 3.5 Parse Targets Complex Documents

Many enterprise documents are difficult to process with simple text extraction. PDFs can contain multiple columns, tables, diagrams, scanned pages and layouts in which information depends on its visual position.

Nemotron 3.5 Parse, a model with 631 million parameters, is designed to extract structured information from these documents so it can be indexed and used by downstream AI applications.

Document parsing is an important part of retrieval because a search system cannot reliably find information that has been extracted incorrectly or lost during preprocessing.

For example, a financial report may contain a revenue figure inside a table, while a technical manual may explain a process through a diagram. A parsing system needs to preserve the relationships between the relevant elements rather than simply extract isolated words.

NVIDIA’s broader retrieval approach combines parsing with embedding and reranking. This allows an enterprise pipeline to process documents, make their contents searchable and prioritize the evidence most relevant to a question.

What the Release Means for Enterprise AI

The new models are aimed at businesses building RAG applications, internal knowledge assistants, document intelligence tools and autonomous agents.

Organizations can use retrieval systems to connect AI applications to company-specific information without requiring every document to be incorporated directly into a model’s training data. This can help applications work with changing information and provide answers grounded in relevant source material.

Open model availability also gives technical teams more control over deployment. Companies can evaluate models against their own data, adapt their pipelines and select infrastructure according to their performance, privacy and cost requirements.

However, open weights do not eliminate deployment costs. Enterprises still need suitable computing resources, storage, indexing systems, monitoring and security controls. They must also evaluate retrieval accuracy, protect sensitive documents and ensure that the final AI application cites or otherwise identifies its evidence appropriately.

Availability and Licensing

NVIDIA says the Nemotron Retriever models are available through Hugging Face, partner platforms and as NVIDIA NIM microservices on build.nvidia.com. The collection is released under the OpenMDW-1.1 license, which NVIDIA presents as supporting commercial use.

Organizations considering deployment should review the applicable license, model cards, hardware requirements and usage documentation before integrating the models into production systems.

The practical benefits will depend on how well the models perform against a company’s own documents and search queries. Public benchmark scores provide a starting point, but testing on representative workloads remains essential.

The Bigger Picture

NVIDIA’s Nemotron 3.5 retrieval release reflects a broader shift in enterprise AI development. As businesses move from chatbots toward agents that perform multi-step tasks, the ability to retrieve accurate evidence becomes increasingly important. A capable language model may still produce unreliable answers if the retrieval layer supplies incomplete, irrelevant or poorly extracted information. Embedding, reranking and parsing address different parts of that problem.

The release also expands NVIDIA’s role beyond supplying chips and computing infrastructure. By offering open models for retrieval and document processing, the company is providing additional components of the software stack used to build AI applications. The long-term value will depend on real-world accuracy, deployment costs and how well these models integrate with existing systems. NVIDIA’s reported benchmark improvements are encouraging, but enterprise teams will need to validate the results against their own requirements before choosing a production configuration.

FAQs

1. What did NVIDIA release in Nemotron 3.5 Retriever?

NVIDIA released five models: three embedding variants, one 8-billion-parameter reranker and one 631-million-parameter document parser.

2. What is the difference between embedding and reranking?

Embedding converts information into numerical representations to help search systems find relevant content. Reranking evaluates the relationship between a query and retrieved candidates, then reorders them by relevance.

3. What is Nemotron 3.5 Parse used for?

It is designed to extract structured information from complex documents, including PDFs, so that their contents can be searched and used by AI applications.

4. How much can the new reranker reduce AI costs?

NVIDIA reported approximately 13%–14% lower agentic token costs per query in selected benchmark configurations. Actual savings will depend on the application, hardware and retrieval pipeline.

5. Are NVIDIA’s Nemotron 3.5 retrieval models available for commercial use?

NVIDIA says the models are openly available for commercial use under the OpenMDW-1.1 license. Developers should review the relevant license and model documentation before deployment.

Looking Ahead

The next test for NVIDIA’s Nemotron 3.5 retrieval models will be their performance in production environments. Businesses will need to assess retrieval quality, document extraction accuracy, latency, memory requirements and total operating costs across their own datasets. Wider adoption could depend on whether the models deliver consistent improvements over existing retrieval systems while remaining straightforward to deploy and maintain.

For the enterprise AI ecosystem, the release adds more open components for building agents that work with proprietary knowledge. If embedding, reranking and parsing can reduce wasted searches and improve evidence quality, they may help organizations build more reliable AI workflows. But the benefits will depend on implementation quality, data governance and the ability to evaluate the complete system rather than relying on individual model benchmarks alone.