🔍 Read the full analysis: How To Use NVIDIA Nemotron 3.5 For Agent Retrieval Tasks on ThorstenMeyerAI.com
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five models for multimodal embedding, reranking and document parsing. The company reports improved retrieval results and lower inference-token use in selected evaluations, but the supplied announcement lacks full methodology, independent validation and detailed information about Parse.
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, targeting AI agents that search enterprise materials such as text, charts and schematics, as detailed in the original analysis. The company says its retrieval setups improved results and reduced agent inference-token use in selected tests, though those results are company-reported and the supplied announcement does not provide independent validation.
The collection comprises Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, a model with 631 million parameters. NVIDIA describes the embedding models as capable of retrieving information from text and visual materials, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to reorder candidate results and help an agent select material for its context.
NVIDIA reports tests on ViDoRe v3, BRIGHT and BrowseComp-Plus in which it held the language model, tools, tasks, corpus and execution policy fixed across compared retrieval configurations. The company says adding Nemotron 3.5 Rerank reduced agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy. It also reports the 8B reranker scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent-token cost.
On ViDoRe v3.1, NVIDIA reports an nDCG@10 score of 61.17 for Embed 8B, compared with 52.23 for Llama Nemotron Embed VL 1B. Embed 1B BF16 scored 58.38, which NVIDIA characterizes as 95.4% of the 8B model’s score. These figures describe the company’s stated evaluation conditions; they do not establish how the models will perform on other collections or agent workloads.
Retrieval Can Affect Agent Cost
Agents often search, inspect results and search again before producing an answer. If retrieval surfaces irrelevant material, the system may spend additional inference tokens processing evidence that does not help resolve the task. Better-ranked results could reduce that repeated work while supplying more relevant evidence, which may matter to organizations operating agents over large document collections.
The multimodal focus addresses another practical retrieval problem: useful information may appear in a chart or table image rather than in text that conventional search can index directly. NVIDIA’s reported results link retrieval quality with agent token use, rather than treating retrieval only as an offline preparation step. However, the real-world effect depends on the document set, language, hardware, workload and integration. Benchmark gains alone do not show that every deployment will use fewer tokens or produce more accurate answers.
NVIDIA Nemotron 3.5 embedding models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Embeddings to Reranking
Embedding models convert queries and documents into representations that search systems can compare for semantic similarity. A reranker takes a set of possible results and scores them against a particular query, helping refine their order. NVIDIA presents the models as components for an agent workflow that searches, inspects evidence and can revise its search in response to what it finds.
The product options vary by size and format. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal alternative. Embed 1B NVFP4 uses quantization for its text and image encoders, with inference efficiency tied to supported hardware. NVIDIA also says its embedding fine-tuning recipes can adapt models to custom collections. In chart and infographic experiments, the company reports relative nDCG@10 gains of 18% for Embed 1B BF16 and 3% for Embed 8B BF16.
NVIDIA says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. The source says support may extend beyond the languages tested, but gives no evaluation results for those additional languages. The announcement positions this release as an extension of the earlier Nemotron 3 Embed text embedding models.
“Adding Nemotron 3.5 Rerank reduces agentic token cost per query by approximately 14% with Nemotron 3.5 Embed 8B, and 13% with Nemotron 3.5 Embed 1B, while improving retrieval accuracy in both configurations.”
— NVIDIA, in its Hugging Face announcement
Benchmark and Parse Details
The supplied announcement does not give a publication date, full benchmark methodology or complete results tables, and it does not include independent replication of the reported scores or token savings. It is also unclear how much the test conditions represent the document collections, tools and agent workflows used by individual organizations.
Details about Nemotron 3.5 Parse are particularly limited: the material gives its parameter count but does not explain its capabilities, supported formats or evaluation. Pricing, specific licensing terms, hardware requirements beyond the NVFP4 deployment note, and a schedule for broader availability are also not provided. The company’s language evaluations cover the stated test sets; performance on other languages and workloads remains unestablished.
Documentation and Independent Tests
The next steps for prospective users are to review the models’ documentation and artifacts for licensing, deployment requirements and instructions for reproducing the reported evaluations. The supplied source describes the models as open and commercially available, but does not provide release links or detailed terms.
Independent tests on enterprise document collections could show whether the reported retrieval gains and token reductions carry over to other corpora and agent tasks. More documentation on Parse would clarify how it works alongside the embedding and reranking models. Until those details are available, the announcement establishes NVIDIA’s reported benchmark results, not a general performance guarantee.
Key Questions
What models did NVIDIA announce?
The collection includes Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, which NVIDIA lists at 631 million parameters.
How are the embedding and reranking models different?
Embedding models help search systems find semantically related material. A reranker scores candidate documents against a query to refine their order before an agent uses the results.
What performance gains did NVIDIA report?
NVIDIA says adding its reranker cut agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy in the tested configurations. These are company-reported results, not independently validated findings in the supplied material.
Can the models retrieve information from images?
NVIDIA describes the embedding models as supporting retrieval from visual material such as tables, charts, infographics, datasheets and schematics, as well as text. Results may vary across document collections and languages.
What is not yet clear about the release?
The supplied announcement does not provide full benchmark methodology, independent test results, detailed licensing and hardware terms, or a full account of Parse’s capabilities and supported formats.
Primary source: Hugging Face · via ThorstenMeyerAI.com
