🔍 Read the full analysis: NVIDIA Nemotron 3.5 Connects Retrieval And Reasoning 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 open models for multimodal embedding, reranking and document parsing. The company reports higher 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 with five open models for multimodal embedding, reranking and document parsing, targeting AI agents that search large collections of text and image-rich files, as detailed in the original analysis. The company says selected evaluations showed improved retrieval accuracy alongside lower agent inference-token use, though the results are company-reported and the supplied announcement does not include independent validation.
The collection comprises Nemotron 3.5 Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse at 631 million parameters. NVIDIA describes the embedding models as finding relevant information in text and visual content, including tables, charts, infographics, datasheets and schematics. A reranker scores query-document pairs to help an agent place the most relevant candidate results higher in its working context.
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 tests, 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. NVIDIA 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 a separate reported result on ViDoRe v3.1, NVIDIA gives Embed 8B an nDCG@10 score of 61.17, compared with 52.23 for Llama Nemotron Embed VL 1B. Embed 1B BF16 scored 58.38, which NVIDIA describes as 95.4% of the 8B model’s score. These figures describe the stated tests; they do not establish performance on other document collections or agent workloads.
Retrieval Can Affect Agent Cost
For an AI agent, retrieval shapes which evidence reaches the language model before it produces an answer. If search repeatedly returns irrelevant material, an agent may spend additional inference tokens inspecting it or searching again. More relevant results could reduce that repeated work while giving the model stronger evidence, a potential benefit for organizations running agents over large document repositories.
NVIDIA’s announcement frames retrieval quality and token use as linked measures rather than treating search as a separate preliminary step. The multimodal focus also addresses a practical document-search problem: information may be embedded in a chart or table image instead of appearing in text that a conventional search system can index. However, real-world savings are not established by these benchmark results; they will depend on the corpus, agent design, hardware and integration choices.
Products Worth Considering
Apple MacBook Air 13.6-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Midnight
Samsung Galaxy Tab A11+ (6GB RAM, 128GB Storage) – Gray
Amazon Kindle Scribe (3rd Generation), 64GB, 11-inch Display, Graphite, with Premium Pen
As an affiliate, we earn on qualifying purchases.
From Embeddings to Reranking
Embedding models encode queries and documents into representations that search systems can compare for semantic relevance. A reranker then evaluates a query against candidate documents to refine their order. NVIDIA describes an agent workflow in which the system searches for evidence, inspects results and may search again based on what it finds. The new models extend the company’s earlier Nemotron 3 Embed text-embedding line toward visual and text-based retrieval.
The model sizes and formats offer different deployment options. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal model. The Embed 1B NVFP4 version uses quantization on text and image encoders for inference efficiency on supported hardware. NVIDIA also reports that fine-tuning recipes for custom collections produced relative nDCG@10 gains of 18% for Embed 1B BF16 and 3% for Embed 8B BF16 in chart and infographic experiments. The company says text retrieval was evaluated across 34 languages, with multilingual visual retrieval tested on MIRACL Vision across 18 languages and on KoViDoRe.
“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 full benchmark methodology, complete results tables or independent replication of the reported scores and token savings. Results may differ across corpora, languages and agent tasks, and the source does not establish that the reported improvements will carry over to enterprise deployments.
Details about Nemotron 3.5 Parse remain limited: the announcement gives its parameter count but little information about capabilities, supported formats or evaluation. The supplied material also does not specify licensing terms, pricing, complete hardware requirements, release links or a schedule for broader availability. NVIDIA says support may extend beyond the languages it evaluated, but reports results only for the stated language sets.
Documentation and Independent Tests
The next steps for users are to consult NVIDIA’s model documentation and artifacts for availability, licensing and deployment requirements, then test the models against their own documents and agent workflows. Independent evaluations on enterprise collections would help determine whether the reported retrieval gains and token reductions transfer beyond NVIDIA’s stated benchmarks. Further documentation on Parse would clarify how the model fits with the embedding and reranking components. The supplied source does not give a date for those updates.
Key Questions
What did NVIDIA announce?
NVIDIA expanded its Nemotron Retriever collection with five open models: three embedding variants, one reranker and Nemotron 3.5 Parse, listed at 631 million parameters.
What does the company say the models improve?
NVIDIA reports better retrieval results and about 13% to 14% fewer agent inference tokens per query when adding its reranker in selected configurations. These are company-reported evaluation results, not a guarantee for other workloads.
Can the embedding models search images as well as text?
NVIDIA describes the models as retrieving information from text and visual material, including charts, tables, infographics, datasheets and schematics. The announcement reports multilingual visual retrieval tests on MIRACL Vision across 18 languages and on KoViDoRe.
What is not yet clear?
The supplied material does not provide independent validation, complete benchmark methodology, detailed Parse capabilities, licensing terms, pricing or a broader availability schedule. Performance on a particular organization’s data also remains to be tested.
Primary source: Hugging Face · via ThorstenMeyerAI.com
