🔍 Read the full analysis: What Embedding, Reranking, And Parsing Mean For NVIDIA Nemotron 3.5 on ThorstenMeyerAI.com
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five open models for multimodal embedding, reranking and document parsing. The company reports improved retrieval scores and lower agent inference-token use in selected evaluations, but the supplied announcement lacks full methodology, independent validation and detail on the Parse model.
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, aimed at helping enterprise AI agents retrieve evidence from text and image-rich documents. The company reports improved retrieval results and lower inference-token use in selected tests, though the original analysis does not include independent validation or full details for its Parse model.
The new collection lists 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 tools for finding relevant material in text and visual content, including tables, charts, infographics, datasheets and schematics. A reranker scores query-document pairs to help prioritize candidate results for an agent.
NVIDIA says it compared retrieval configurations on ViDoRe v3, BRIGHT and BrowseComp-Plus, holding the language model, tools, tasks, corpus and execution policy fixed. In those evaluations, 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 describes as retaining 95.4% of the 8B model’s score. These figures are company-reported benchmark results; they do not establish how the models will perform on other data collections or workloads.
Retrieval Can Shape Agent Costs
For agents that search large document collections, retrieval affects both the evidence presented to a language model and the amount of work it must do to find that evidence. If relevant passages or visual details are surfaced earlier, an agent may need fewer repeated searches or less inspection of irrelevant material. NVIDIA’s reported results link retrieval accuracy with agent task performance and token use, rather than treating search as a separate preliminary step.
The multimodal emphasis addresses a practical gap in document-heavy work: information may appear inside a chart or table image rather than in text that conventional search can index. Embedding models intended to retrieve visual material could help agents locate that evidence. The actual effect will depend on factors including the document collection, hardware, workload and system integration; the benchmark results alone do not show that every deployment will save tokens or improve answers.
AI embedding models for enterprise search
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Retriever Models Fit
Embedding models represent queries and documents so a search system can find items that are semantically related, even when they use different wording. A reranker then compares retrieved candidates against the query and reorders them, helping an agent place more relevant material in its working context. NVIDIA describes a workflow in which an agent searches, inspects evidence and may refine its search based on what it finds.
The three embedding options offer different model sizes and formats. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal option. Embed 1B NVFP4 uses quantization for the text and image encoders, which the company presents as an inference-efficiency option for supported hardware. NVIDIA also says its embedding fine-tuning recipes can adapt models to custom collections. In chart and infographic experiments, it 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, as well as on KoViDoRe. The supplied material does not provide the full results tables or establish performance in languages beyond those named evaluation sets.
“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 retrieval tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark and Parse Details Missing
The supplied announcement does not include full benchmark methodology, detailed results tables or independent replication of the reported scores and token savings. It says key parts of the evaluation setup were held fixed, but the information provided is not enough to determine how the results might vary across different corpora, hardware, agent designs or tasks.
The source names Nemotron 3.5 Parse and its parameter count but gives little information about its capabilities, supported formats, evaluation or role alongside the embedding and reranking models. It describes the collection as open and commercially available, but the supplied material does not provide licensing terms, release links, pricing, broader hardware requirements or a schedule for further availability details. NVIDIA also says support may extend beyond tested languages, but does not report results for those additional languages.
As an affiliate, we earn on qualifying purchases.
Documentation and Replication to Follow
Readers and developers will need the model artifacts and documentation to check the models’ availability, licensing and deployment requirements and to assess the Parse model’s intended use. Independent testing on enterprise document collections would help show whether the reported retrieval improvements and token reductions carry over to different workloads. The supplied source does not give a date for those materials or identify a forthcoming release milestone.
AI reranking models for document analysis
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did NVIDIA add to Nemotron Retriever?
NVIDIA announced five models: Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse. Parse is listed as a 631-million-parameter model.
What do the embedding and reranking models do?
Embedding models help search systems find semantically relevant text and visual material. A reranker scores query-document pairs and reorders candidate results to help an agent prioritize useful evidence.
How much lower did NVIDIA say token use was?
In its stated evaluations, NVIDIA reported about 14% fewer agent inference tokens per query when Rerank was used with Embed 8B and about 13% fewer with Embed 1B, alongside improved retrieval accuracy. These are company-reported results, not a guarantee for other workloads.
What is known about Nemotron 3.5 Parse?
The supplied announcement identifies Parse as a 631-million-parameter model, but gives little detail about its capabilities, supported document formats, evaluation or availability.
Have the reported results been independently verified?
The source material does not cite independent replication. It describes NVIDIA’s evaluation setup and results, but does not provide enough detail to establish how those outcomes generalize to other systems and document collections.
Primary source: Hugging Face · via ThorstenMeyerAI.com
