🔍 Read the full analysis: NVIDIA Nemotron 3.5 Brings Embedding, Reranking, And Parsing To AI Agents 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 scores and lower inference-token use in selected evaluations, but the results have not been independently validated and details about the Parse model remain limited.
NVIDIA has expanded its Nemotron Retriever collection with five open models for multimodal embedding, reranking and document parsing, targeting AI agents that search business documents containing text, charts and other visual material. The company says selected evaluations showed better retrieval results alongside lower agent inference-token use, though the supplied announcement does not include independent validation or full details for every model.
The five models are 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 matching queries to relevant material in text and visual content, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize results for its context.
NVIDIA reports that 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. In a comparison on ViDoRe v3.1, it reports an nDCG@10 score of 61.17 for Embed 8B, against 52.23 for Llama Nemotron Embed VL 1B. Embed 1B BF16 scored 58.38, which NVIDIA says represents 95.4% of the 8B model’s score.
The company also says its Embed 8B reranker configuration scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent token cost. Its evaluations included ViDoRe v3, BRIGHT and BrowseComp-Plus. NVIDIA says it held the language model, tools, tasks, corpus and execution policy fixed while comparing retrieval setups. These figures describe the reported tests, not guaranteed outcomes on other systems or document collections.
Retrieval Can Affect Agent Costs
Retrieval quality can shape both the evidence an AI agent uses and how much work it performs before producing an answer. When an agent repeatedly searches, receives irrelevant material and tries again, it may consume more inference tokens per query. NVIDIA’s results connect retrieval performance with token use in selected tests, making the models relevant to businesses evaluating the operating cost and answer quality of document-based agents.
The multimodal focus addresses another practical issue: important information may be embedded in a chart or table image rather than available as searchable text. Models that retrieve visual evidence could help agents handle questions spanning different document formats. Whether this produces measurable benefits will depend on the collection, workload, hardware and integration; the announcement does not establish that the reported gains will carry over to every enterprise deployment.
As an affiliate, we earn on qualifying purchases.
From Text Search to Visual Documents
Embedding models represent queries and documents in a form that supports semantic search, while rerankers compare candidate results with a query to reorder them. NVIDIA presents the models as part of an agent workflow in which the system searches for evidence, inspects results and may search again. The new collection extends the company’s earlier Nemotron 3 Embed text embedding models toward visual and document-rich retrieval.
The lineup offers 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 text and image encoders, which NVIDIA says can support inference efficiency on compatible hardware. The company also reports that fine-tuning recipes improved chart and infographic retrieval: relative nDCG@10 gains of 18% for Embed 1B BF16 and 3% for Embed 8B BF16 in its experiments.
NVIDIA says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. Those are the language ranges specified for the reported evaluations; the announcement’s suggestion that support may extend further does not establish performance in additional languages.
“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
As an affiliate, we earn on qualifying purchases.
Benchmark and Parse Details
The supplied announcement does not provide full benchmark methodology, complete results tables or independent replication of the reported scores and token savings. Results on other corpora, languages and agent tasks may differ, and the source does not quantify how performance changes across hardware or deployment settings.
Details about Nemotron 3.5 Parse are especially limited: the source gives its parameter count but does not fully describe its capabilities, supported formats or evaluation. It also does not specify licensing terms, pricing, release links or hardware requirements beyond noting the NVFP4 deployment consideration. The announcement describes the models as open and commercially available, but the supplied material does not clarify the terms or timing of broader availability.
As an affiliate, we earn on qualifying purchases.
Documentation and Independent Tests
Model documentation and artifacts will be needed to clarify deployment requirements, licensing and the capabilities of Parse. Users and researchers can then test whether NVIDIA’s reported retrieval improvements and token savings hold on different enterprise collections and agent workloads. The supplied source gives no schedule for further releases or updates, so a timeline for additional technical details is not confirmed.
AI reranking models for enterprise
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did NVIDIA announce?
NVIDIA expanded its Nemotron Retriever collection with five models for embedding, reranking and document parsing, intended to support agents searching text and visual material.
What are the five models?
The collection includes Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, a 631-million-parameter model.
What performance gains did NVIDIA report?
NVIDIA reports lower agent inference-token use—about 14% with Embed 8B and 13% with Embed 1B when paired with its reranker—alongside improved retrieval accuracy in selected tests. These are company-reported benchmark results, not guarantees for other workloads.
Have the results been independently verified?
The supplied announcement does not include independent validation or full benchmark methodology and results tables. Performance on other collections and tasks remains to be tested.
What is not yet known about Parse?
The source identifies Parse as a 631-million-parameter model but does not fully explain its capabilities, supported document formats, evaluation results or deployment requirements.
Primary source: Hugging Face · via ThorstenMeyerAI.com
