🔍 Read the full analysis: Building AI Agents With NVIDIA Nemotron 3.5 Retrieval Models on ThorstenMeyerAI.com
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five models for embedding, reranking and document parsing, aimed at helping AI agents retrieve evidence from text and visual documents. The company reports better retrieval results and lower agent inference-token use in selected evaluations, but the supplied announcement lacks full methods, 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 need to find evidence across text and image-rich documents. The company says selected evaluations showed improved retrieval results and lower agent inference-token use, though the figures are NVIDIA-reported and the supplied announcement does not include independent validation or complete methodology.
The collection includes Nemotron 3.5 Embed 8B BF16, Embed 1B BF16 and Embed 1B NVFP4, as well as Rerank 8B BF16 and Parse, a 631-million-parameter model. NVIDIA positions the embedding models for retrieving information from text and visual material, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs to help an agent prioritize candidate results in the context it uses to form an answer.
In evaluations involving 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 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. These are results from the stated test setups, not guaranteed savings for other systems or workloads.
For a separate ViDoRe v3.1 comparison, 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; NVIDIA says that is 95.4% of the 8B model’s score. The source also reports relative nDCG@10 gains in chart and infographic experiments after fine-tuning: 18% for Embed 1B BF16 and 3% for Embed 8B BF16. The source does not provide full results tables for these evaluations.
Retrieval Shapes Agent Costs
Search quality can affect both the evidence an agent receives and the amount of work it performs. If an agent repeatedly retrieves irrelevant documents or misses useful passages, it may need more searches and consume more inference tokens before producing an answer. NVIDIA’s reported results frame retrieval quality and token use as connected measures, rather than treating search as a separate step from agent performance.
The multimodal focus addresses a practical challenge for organizations with document-heavy workflows: relevant information may appear in a chart, diagram or scanned page rather than in text that conventional search can readily index. Models that retrieve such visual evidence could help agents handle questions spanning different document formats. Whether they do so reliably will depend on the collection, task, hardware and integration. The benchmark results alone do not establish savings or accuracy gains in a particular company’s environment.
NVIDIA Nemotron 3.5 retrieval models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Text Embeddings to Visual Search
NVIDIA presents Nemotron 3.5 Embed as an extension of its earlier Nemotron 3 Embed text models. Embedding models represent queries and documents so a search system can identify semantically related material; a reranker then scores candidate results against the query to refine their order. In the workflow described by NVIDIA, an agent searches for evidence, inspects what it finds and may search again based on that evidence.
The model sizes and formats offer different deployment choices. NVIDIA positions Embed 8B BF16 for quality-focused retrieval and Embed 1B BF16 as a smaller multimodal option. Embed 1B NVFP4 applies quantization to text and image encoders for inference efficiency on supported hardware. The company also provides embedding fine-tuning recipes intended to adapt models to custom document collections, though results from the reported chart and infographic experiments may not transfer to other data.
NVIDIA says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. Those evaluation sets establish the scope of the results reported in the source; they do not confirm performance for additional languages or every document type.
“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 Scope and Parse Details
The supplied announcement does not state its publication date, provide complete benchmark methods or results tables, or cite independent replication of the reported scores and token savings. It also gives little information about Nemotron 3.5 Parse beyond its name and 631-million-parameter count; supported formats, capabilities, evaluation results and its role alongside the other models remain unclear.
The source describes the collection as open and commercially available, but does not provide licensing terms, release links, detailed pricing or a schedule for broader availability. Hardware requirements are also not fully specified, apart from the note that the NVFP4 option is intended for supported hardware. Performance may vary across document collections, languages and agent tasks, and the announcement does not establish results beyond the evaluations it names.
AI agent document parsing software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Documentation and Independent Testing
Readers seeking to assess the models will need the model artifacts and documentation for release details, licensing, supported hardware and deployment instructions. The supplied source does not give a timeline for further announcements or state when more information about Parse will be published.
Independent tests on enterprise document collections would help determine whether NVIDIA’s reported retrieval gains and token reductions extend to different workloads. Comparisons that disclose their data, methods and compute setup could also help organizations judge trade-offs among model size, retrieval quality and inference cost. Until such evidence is available, the reported figures should be treated as results from NVIDIA’s stated evaluations, not as a forecast for every deployment.
visual and text document retrieval tools
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: Embed 8B BF16, Embed 1B BF16, Embed 1B NVFP4, Rerank 8B BF16 and Parse at 631 million parameters.
What do the embedding and reranking models do?
The embedding models help search systems retrieve semantically related text and visual material. The reranker scores query-document pairs to help prioritize which results an agent uses as evidence.
What performance gains does NVIDIA report?
NVIDIA says adding its reranker reduced agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B, while improving retrieval accuracy in its selected tests. The source also reports ViDoRe v3.1 nDCG@10 scores of 61.17 for Embed 8B and 58.38 for Embed 1B BF16. These results have not been independently validated in the supplied material.
Is Nemotron 3.5 Parse fully documented in the announcement?
No. The supplied source identifies Parse as a 631-million-parameter model but does not give full details on its functions, supported formats, availability or evaluation results.
Do the reported results guarantee gains for other deployments?
No. The figures apply to NVIDIA’s stated evaluations. Results in other settings may differ based on the document collection, language, task, hardware and integration.
Primary source: Hugging Face · via ThorstenMeyerAI.com
