🔍 Read the full analysis: From Embedding To Parsing: NVIDIA Nemotron 3.5 For AI Agents on ThorstenMeyerAI.com
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing. The company reports improved retrieval scores and 13% to 14% lower agent inference-token use in selected evaluations, but the supplied announcement lacks independent validation and detailed information about the Parse model.
NVIDIA has expanded its Nemotron Retriever collection, as detailed in the original analysis, with five open models for multimodal embedding, reranking and document parsing, targeting AI agents that search text and visual material in enterprise documents. The company says selected tests showed improved retrieval accuracy alongside lower agent inference-token use, though the reported results have not been independently validated in the supplied material.
The collection comprises Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, with 631 million parameters. NVIDIA positions the embedding models for finding relevant information in text and visual content, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize candidate results.
NVIDIA says it compared retrieval configurations on ViDoRe v3, BRIGHT and BrowseComp-Plus, holding the language model, tools, tasks, corpus and execution policy fixed. It 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. The source does not provide full methodology or independent replication.
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 the company describes as 95.4% of the 8B model’s score. These are results from the stated evaluations, not guarantees for other collections or agent workloads.
Retrieval Could Cut Agent Work
For agents that repeatedly search large document collections, retrieval quality can affect both the evidence used to answer a question and how much model inference is required. If the reported results carry over to real deployments, more relevant search results could reduce repeated inspection of irrelevant material and lower token use per task. The potential benefit is a more efficient agent workflow, not simply a higher standalone search score.
The multimodal focus addresses a specific challenge in document-heavy work: useful information may appear in a chart, table or schematic rather than in text that conventional search can readily match. Embedding models designed to retrieve visual material could help agents locate that evidence. Actual results will depend on the document collection, model integration, hardware and task design; NVIDIA’s benchmarks do not establish performance across all enterprise settings.
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 Parsing
NVIDIA presents Nemotron 3.5 Embed as an extension of its earlier Nemotron 3 Embed text embedding models. Embeddings represent queries and documents in a form that lets search systems identify semantically related material. A reranker then evaluates candidate query-document pairs to refine their order. In the workflow described by NVIDIA, an agent searches, inspects evidence and may search again based on what it finds.
The models offer different deployment choices: Embed 8B BF16 is positioned for quality-focused retrieval, while the 1B BF16 model offers a smaller multimodal option. The 1B NVFP4 version applies quantization to text and image encoders for inference efficiency on supported hardware. NVIDIA also reports that fine-tuning recipes improved nDCG@10 in chart and infographic experiments by 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. The announcement says support may extend beyond tested languages, but does not provide evaluation results for additional languages. The Parse model is named as part of the collection, but the supplied information does not explain its role in the workflow in comparable detail.
“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 Gaps
The supplied announcement does not provide a publication date, full benchmark methodology, complete results tables or independent replication of the reported scores and token savings. The comparisons are tied to particular evaluation setups, and performance on other corpora, languages and agent tasks remains unknown.
Details about Nemotron 3.5 Parse are also limited to its name and 631 million parameters. The source does not specify its capabilities, supported document formats or evaluation results. It also does not provide licensing terms, pricing, full hardware requirements or release links, so those details cannot be confirmed from the supplied material.
Documentation and Independent Tests
The next useful step is publication or review of the model artifacts and documentation, which could clarify availability, licensing, hardware needs and how Parse fits alongside embedding and reranking. Independent testing on enterprise document collections would help determine whether NVIDIA’s reported retrieval gains and token reductions transfer to different workloads.
No further release date or specific follow-up milestone is given in the source. Until more documentation and testing are available, the benchmark figures should be treated as company-reported results for the stated evaluations rather than as expected outcomes for every deployment.
Key Questions
What models did NVIDIA add to the Nemotron Retriever collection?
The five listed models are Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse. NVIDIA identifies Parse as a 631-million-parameter model.
What token savings did NVIDIA report?
NVIDIA reports about 14% fewer agent inference tokens per query when adding Rerank to Embed 8B, and about 13% with Embed 1B. These figures apply to the company’s stated evaluation setup and have not been independently validated in the supplied material.
Can the embedding models search visual information?
NVIDIA says the embedding models are designed to retrieve information from text and visual material, including tables, charts, infographics, datasheets and schematics. Results in other document collections have not been established by the supplied announcement.
What is known about Nemotron 3.5 Parse?
The source names Parse as a 631-million-parameter model, but does not describe its supported formats, capabilities, availability or evaluation results in detail.
Are the reported benchmark results independently confirmed?
No independent replication is included in the supplied source. NVIDIA reports results from specified benchmarks, but the methodology and full results tables are not provided there.
Primary source: Hugging Face · via ThorstenMeyerAI.com
