🔍 Read the full analysis: A Closer Look At NVIDIA Nemotron 3.5’S Retrieval Models For Agents on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five models for embedding, reranking and document parsing, including models intended to retrieve evidence from text and visual material. The company reports improved retrieval scores and lower inference-token use in selected evaluations, but the supplied announcement does not provide full methodology, independent validation or meaningful details about the Parse model.
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, targeting AI agents that search across text and image-rich documents. The company says the retrieval models improved results and reduced agent inference-token use in selected benchmarks, though the reported gains have not been independently validated in the supplied material.
The collection includes Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, with 631 million parameters. NVIDIA describes the embedding models as tools for retrieving information from text and visual content, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize candidate evidence.
In evaluations involving ViDoRe v3, BRIGHT and BrowseComp-Plus, NVIDIA says it held the language model, tools, tasks, corpus and execution policy fixed while comparing retrieval setups. The company reports that 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 says Embed 8B scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent token cost.
For 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 says is 95.4% of the 8B model’s score. These are company-reported results tied to the stated evaluations, not a guarantee of performance on other collections or tasks.
Retrieval Can Shape Agent Costs
For an agent working through a large document collection, retrieval quality can affect both the evidence it receives and the amount of model work required. If a system repeatedly brings back irrelevant material, the agent may need more searches and more inference tokens before it can answer. NVIDIA’s reported results link retrieval accuracy with token use, rather than treating search as a step separate from agent performance.
The multimodal emphasis addresses documents where key information appears in a chart, schematic or table image rather than in searchable text. A model that retrieves those details could be useful for document-heavy business workflows. But actual effects depend on the documents, hardware, workload and integration; benchmark scores alone do not establish cost savings or accuracy gains for a particular company.
NVIDIA Nemotron 3.5 retrieval models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Text Search to Visual Documents
NVIDIA presents Nemotron 3.5 Embed as an extension of its earlier Nemotron 3 Embed text models. Embeddings represent queries and documents in a form that supports semantic search; a reranker then compares candidate documents with the query to refine their order. In NVIDIA’s described agent workflow, the system searches for evidence, inspects results and may search again based on what it finds.
The models offer different size and deployment options. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal choice. The Embed 1B NVFP4 version uses quantization in its text and image encoders for inference efficiency on supported hardware. NVIDIA also reports that fine-tuning recipes for custom collections produced relative nDCG@10 gains in chart and infographic experiments: 18% for Embed 1B BF16 and 3% for Embed 8B BF16.
The company says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. Those figures describe the reported evaluation coverage; the source does not establish results for other languages it says may be supported.
“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
multimodal document embedding tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark Scope and Parse Details
The supplied announcement does not provide full benchmark methodology or results tables, and it does not include independent validation of the reported scores or token savings. Results could differ on other corpora, languages, agent tasks and deployment setups. The company’s description of fixed evaluation conditions helps frame its comparisons, but does not by itself show how well the models will perform in a specific enterprise environment.
Details about Nemotron 3.5 Parse are particularly limited: the source names its parameter count but does not explain its capabilities, supported formats, evaluation or how it works alongside the embedding and reranking models. The supplied material also does not state licensing terms, pricing, specific hardware requirements beyond the NVFP4 note, or the timing and links for broader availability.
AI agent document parsing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Documentation and Independent Tests
The next useful step for prospective users is to review the model artifacts and documentation for availability, licensing and deployment requirements, then test the models against their own document collections. Independent evaluations could show whether the reported retrieval gains and token reductions hold under different workloads. Further information about Parse would clarify its role in the collection; the supplied announcement does not give a schedule for additional details or updates.
visual and text retrieval AI tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did NVIDIA add to the Nemotron Retriever collection?
NVIDIA lists five models: Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, a 631-million-parameter model.
What does the reranker do?
NVIDIA describes the reranker as scoring query-document pairs so an agent can prioritize more relevant search results for its context.
How much lower was agent token use in NVIDIA’s tests?
NVIDIA reports about 14% fewer inference tokens per query with Embed 8B and 13% fewer with Embed 1B when using Nemotron 3.5 Rerank, alongside improved retrieval accuracy. These are company-reported outcomes for selected evaluations.
Have the benchmark claims been independently confirmed?
The supplied source does not report independent replication. It also lacks full methodology and results tables, so the claims should be read as NVIDIA’s reported results rather than general performance guarantees.
What is known about Nemotron 3.5 Parse?
The source identifies Parse as a 631-million-parameter model, but does not provide enough information about its capabilities, supported document formats, evaluation or availability to explain its practical role.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
