🔍 Read the full analysis: Can NVIDIA Nemotron 3.5 Retrieval Models Help AI Agents Find Better Answers? 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 tests, but the supplied announcement lacks full methodology and independent validation.
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, as detailed in the original analysis, targeting AI agents that search text and image-rich documents. The company says selected evaluations showed improved retrieval accuracy and lower inference-token use, though the results have not been independently validated in the supplied material.
The new collection includes Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse, a 631-million-parameter model. NVIDIA describes the embedding models as finding relevant material across text and visual content, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help prioritize candidate results for an agent.
In evaluations involving ViDoRe v3, BRIGHT and BrowseComp-Plus, NVIDIA says it kept the language model, tools, tasks, corpus and execution policy fixed while comparing retrieval configurations. 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. Those figures describe the tested setups, not expected savings for every agent or document collection.
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 NVIDIA characterizes as 95.4% of the 8B model’s score. In another comparison, the company says its 8B reranker scored about 0.9 points higher in nDCG@10 than Qwen3-VL-Reranker-8B at a similar agent token cost.
Retrieval Quality Can Cut Agent Work
For agents that search large document collections, retrieval affects both the evidence available to answer a question and the amount of work needed to find it. If an agent repeatedly retrieves irrelevant passages or misses information embedded in charts, it may need additional searches and consume more inference tokens. Better retrieval could reduce that repeated effort while giving the model more relevant evidence.
NVIDIA’s announcement presents the models as a way to improve both retrieval quality and token efficiency, rather than treating search as a separate preprocessing step. The multimodal focus addresses documents where important details may appear in a visual element instead of text that conventional search can readily index. Whether that helps in practice will depend on the corpus, language, hardware, agent design and integration choices.
The reported results may give developers benchmarks to examine, but they do not establish cost savings or answer quality for a particular business workflow. The announcement’s lack of independent replication and full evaluation details means organizations would need to test the models against their own documents and tasks before drawing conclusions.
Products Worth Considering
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
How the Model Set Fits Together
Embedding models turn queries and documents into representations that search systems can compare for semantic relevance. A reranker then scores candidate documents against a query, helping an agent choose which results to inspect or include as evidence. NVIDIA describes a workflow in which an agent searches, examines retrieved material and may search again based on what it finds.
The collection offers several deployment choices. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal option. Embed 1B NVFP4 uses quantization for its text and image encoders, which the company describes as an inference-efficiency option on supported hardware. NVIDIA also says its 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 text retrieval was evaluated across 34 languages, while multilingual visual retrieval was tested on MIRACL Vision across 18 languages and on KoViDoRe. The announcement suggests support may extend beyond the tested language sets, but does not provide evaluation results for those additional languages. It also names the Parse model, but gives little information about its functions or testing.
“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 Release Details
The supplied announcement does not include a publication date, full benchmark methodology or complete results tables, and it provides no independent replication of the reported scores or token savings. The reported outcomes may differ on other document collections, languages, hardware and agent tasks. NVIDIA’s comparisons also do not establish performance across every possible retrieval setup.
Details about Nemotron 3.5 Parse remain particularly limited: the source identifies its parameter count but does not explain supported document formats, capabilities or evaluation results. The supplied material also does not spell out licensing terms, pricing, hardware requirements beyond the NVFP4 note, release links or a schedule for broader availability. Although the models are described as open and commercially available, those details could not be confirmed from the provided source.
Documentation and Independent Tests
The next step for prospective users is to review NVIDIA’s model artifacts and documentation for licensing, deployment requirements and evaluation procedures. Independent tests on enterprise document collections could show whether the reported retrieval gains and token reductions carry over to different workloads.
More information about Parse would clarify how it works alongside the embedding and reranking models, while detailed release terms would help developers assess whether and how they can deploy the collection. Until those details and external evaluations are available, the benchmark figures remain NVIDIA-reported results from specified tests.
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, which NVIDIA lists at 631 million parameters.
What token savings does NVIDIA report?
NVIDIA says adding Nemotron 3.5 Rerank cut agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B in its selected evaluations, while improving retrieval accuracy. These are company-reported results, not a guarantee for other workloads.
Can the embedding models retrieve information from charts and images?
NVIDIA describes the embedding models as retrieving information from visual material such as tables, charts, infographics, datasheets and schematics, alongside text. The supplied source does not establish how well they perform on every document type or collection.
Have the benchmark results been independently verified?
The supplied announcement does not include independent replication. NVIDIA reports results from evaluations including ViDoRe v3, BRIGHT and BrowseComp-Plus, but full methodology and results tables are not provided in the source material.
What is known about Nemotron 3.5 Parse?
The source identifies Parse as a 631-million-parameter model, but provides little detail about its capabilities, supported formats, availability or evaluation results.
Primary source: Hugging Face · via ThorstenMeyerAI.com
