🔍 Read the full analysis: NVIDIA Nemotron 3.5 Offers New Retrieval Tools For AI 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 multimodal embedding, reranking and document parsing. The company reports better retrieval scores and 13% to 14% fewer agent inference tokens per query in selected evaluations, but the supplied announcement lacks full methodology, independent validation and 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 text and image-rich documents. The company says configurations using the new models improved retrieval results and reduced agent inference-token use in selected tests, though the supplied announcement does not provide independent validation or full benchmark details.
The collection includes Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Nemotron 3.5 Parse, which NVIDIA lists at 631 million parameters. NVIDIA describes the embedding models as tools for retrieving relevant information from text and visual material, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize results for its context.
In evaluations using ViDoRe v3, BRIGHT and BrowseComp-Plus, NVIDIA says it held the language model, tools, tasks, corpus and execution policy fixed while comparing retrieval configurations. It reports that adding Rerank 8B reduced agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy. These are NVIDIA-reported outcomes under the stated test setup, not guarantees for other collections or workloads.
NVIDIA also reports that Embed 8B scored 61.17 nDCG@10 on ViDoRe v3.1, compared with 52.23 for Llama Nemotron Embed VL 1B. Embed 1B BF16 scored 58.38, which the company says is 95.4% of the 8B model’s score. In a separate comparison, NVIDIA says Embed 8B with its reranker scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent token cost. Benchmark results reflect specific datasets and methods; performance elsewhere may differ.
Retrieval Can Shape Agent Costs
For an agent working across a large document collection, retrieval determines which evidence reaches the language model. If search returns irrelevant material, the agent may need more searches or repeated inspection before answering. NVIDIA’s reported results link retrieval quality with both task performance and inference-token use, making search and reranking relevant to the operating costs of agent systems, not just a separate indexing step.
The visual retrieval focus addresses a common gap in document-heavy work: a needed fact may be embedded in a chart, table or schematic rather than available as searchable text. Models that can locate such evidence could make it easier for agents to answer questions spanning different document formats. The practical effect will depend on the quality of the source documents, the collection, hardware and how each model is integrated. The reported benchmarks do not establish savings or accuracy gains for every deployment.
Products Worth Considering

Samsung Galaxy Tab S10 Lite, 10.9-inch LCD, 6GB RAM, 128GB Storage, Gray

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
As an affiliate, we earn on qualifying purchases.
From Embeddings to Visual Search
Embedding models represent queries and documents in a form that lets a search system identify semantically related material. A reranker then scores candidate documents against a query to refine their order. NVIDIA presents Nemotron 3.5 as supporting an agent workflow in which the system searches, reviews evidence and may search again based on what it finds. The company says the new embedding models extend its earlier Nemotron 3 Embed text models with multimodal retrieval capabilities.
The lineup offers different size and deployment options: NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal choice. Embed 1B NVFP4 uses quantization for its text and image encoders, which NVIDIA presents as an inference-efficiency option on supported hardware. The company also says its fine-tuning recipes can adapt embedding 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; those gains are tied to the company’s stated experiments.
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 source says support may extend beyond the languages tested, but does not report benchmark results for those additional languages. The collection’s models are described as open and commercially available, but the supplied announcement does not include release links or licensing terms.
“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 Missing
The supplied announcement does not give a publication date, complete benchmark methodology or full results tables, and it does not include independent replication of the reported scores or token savings. The results are company-reported and may not carry over to other document collections, languages, agent tasks or hardware setups. The source also does not spell out the evaluation conditions in enough detail to independently assess each comparison.
Details for Nemotron 3.5 Parse are especially limited: the announcement gives its parameter count but does not describe its capabilities, supported formats, availability or evaluation results. Pricing, licensing terms, specific hardware requirements beyond the NVFP4 note, and the timing of broader availability are also absent from the supplied material. NVIDIA’s statements about potential language support beyond the evaluated sets are not equivalent to reported results in those languages.
Documentation and Independent Tests
Users and developers will need the model artifacts and documentation to check availability, licensing, deployment requirements and how to reproduce NVIDIA’s evaluations. The supplied source does not specify a release schedule for updates or additional technical material. Independent testing on enterprise document collections could show whether the reported retrieval gains and token reductions hold across different data, hardware and agent workflows. More documentation on Parse would clarify how it fits alongside the embedding and reranking models.
Key Questions
What did NVIDIA announce?
NVIDIA expanded its Nemotron Retriever collection with five models: three embedding variants, one reranker and Nemotron 3.5 Parse, listed at 631 million parameters.
What do the embedding and reranking models do?
The embedding models help search systems retrieve relevant text and visual material. The reranker scores query-document pairs to help prioritize results for an agent’s context.
How much did NVIDIA say token use fell?
NVIDIA reports about 14% fewer agent inference tokens per query with Embed 8B and Rerank 8B, and about 13% with Embed 1B and Rerank 8B, alongside improved retrieval accuracy in the tested configurations.
Have the benchmark results been independently verified?
The supplied announcement reports NVIDIA’s own results and does not provide independent validation. The figures are tied to selected tests and do not establish performance on other workloads.
What is known about Nemotron 3.5 Parse?
The announcement identifies Parse as a 631-million-parameter model, but the supplied material does not give full details about its functions, supported formats, availability or evaluation.
Primary source: Hugging Face · via ThorstenMeyerAI.com
