🔍 Read the full analysis: NVIDIA Nemotron 3.5: Three Retrieval Models For More Capable AI Agents on ThorstenMeyerAI.com
TL;DR
NVIDIA has expanded its Nemotron Retriever collection with five open models for multimodal embedding, reranking and document parsing. The company reports better retrieval scores and lower agent inference-token use in selected evaluations, but the supplied announcement lacks independent validation and detailed information about Parse.
NVIDIA has expanded its Nemotron Retriever collection with five open models for embedding, reranking and document parsing, targeting AI agents that need to retrieve evidence from text and image-rich documents, as detailed in the original analysis. The company reports that selected retrieval setups improved benchmark results while reducing inference tokens used per query, but those figures come from NVIDIA’s evaluations and have not been independently validated in the supplied material.
The announced models are 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 retrieving information from text and visual content, including tables, charts, infographics, datasheets and schematics. The reranker scores query-document pairs so an agent can prioritize candidate results for its context.
NVIDIA says its comparisons on ViDoRe v3, BRIGHT and BrowseComp-Plus held 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.
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; NVIDIA describes that as 95.4% of the 8B model’s score. These are company-reported outcomes for particular tests, not a guarantee of performance on other workloads.
Retrieval Can Cut Agent Work
Retrieval affects both the evidence an agent receives and the work it performs to find that evidence. If an agent repeatedly searches irrelevant material, it may spend more inference tokens before producing an answer. More relevant results could reduce that repeated work while improving the material available to the language model. NVIDIA’s reported results connect retrieval accuracy with token use in selected agent tests, rather than measuring search only as a separate preprocessing step.
The multimodal focus addresses a practical problem for organizations with document-heavy workflows: information may be embedded in a chart, table or diagram rather than captured in searchable text. Models that can retrieve such material could help agents answer questions requiring evidence across formats. Whether this changes cost or answer quality in practice will depend on the document collection, hardware, workload and integration; the reported benchmarks do not establish the same results for every enterprise deployment.
NVIDIA Nemotron 3.5 retrieval models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Embeddings to Reranking
Embedding models represent queries and documents in a form that allows a search system to find semantically related material. A reranker then scores candidate results against a query and can reorder them before an agent uses them. NVIDIA presents Nemotron 3.5 as an extension of its earlier Nemotron 3 Embed text models, adding visual retrieval and a reranking model to the collection.
The model sizes offer different deployment choices, according to NVIDIA. Embed 8B BF16 is positioned for retrieval quality, while Embed 1B BF16 is a smaller multimodal option. Embed 1B NVFP4 uses quantization on text and image encoders for inference efficiency on supported hardware. NVIDIA 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.
NVIDIA says it evaluated text retrieval across 34 languages, and multilingual visual retrieval on MIRACL Vision across 18 languages and on KoViDoRe. The source does not provide full results for these evaluations or establish performance in languages beyond those tested.
“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 and Parse Gaps
The supplied announcement does not include full benchmark methodology, complete results tables or independent replication of the reported scores and token savings. It is also unclear how the models perform on different corpora, agent tasks, languages and hardware configurations. NVIDIA’s figures should be read as results from its stated tests, not as established performance across general enterprise use.
Details about Nemotron 3.5 Parse are particularly limited: the source gives its parameter count but does not explain its capabilities, supported document formats, evaluation or relationship to the embedding and reranking models. The material also does not provide licensing terms, pricing, release links, specific hardware requirements beyond the NVFP4 note, or a schedule for broader availability. Although NVIDIA describes the models as open and commercially available, the supplied source does not establish those practical release details.
As an affiliate, we earn on qualifying purchases.
Documentation and Replication
The next useful step for adopters is to review the model artifacts and documentation for availability, licensing and deployment requirements, then test the models against their own documents and agent tasks. Independent evaluations would help show whether the reported retrieval gains and token reductions carry over to different collections and operating conditions.
Further documentation on Parse would clarify its intended role in the collection, while fuller benchmark disclosures could make NVIDIA’s comparisons easier to reproduce. The supplied announcement does not give a date for those updates, so their timing remains unknown.
As an affiliate, we earn on qualifying purchases.
Key Questions
What did NVIDIA announce?
NVIDIA announced five open Nemotron 3.5 retrieval models: three embedding variants, one reranker and Parse, a 631-million-parameter model. They are intended for retrieval across text and image-rich documents.
What performance gains did NVIDIA report?
NVIDIA says adding its reranker cut agent inference tokens per query by about 14% with Embed 8B and 13% with Embed 1B while improving retrieval accuracy in its selected evaluations. It also reported model scores on ViDoRe tests. These results have not been independently validated in the supplied material.
What does the reranker do?
A reranker scores candidate documents against a query and helps reorder them so an agent can use more relevant results. NVIDIA says this improved retrieval accuracy and lowered token use in its stated test configurations.
What is known about Nemotron 3.5 Parse?
The supplied announcement identifies Parse as a 631-million-parameter model, but does not detail its capabilities, supported formats or evaluation. Its role relative to the embedding and reranking models remains unclear.
Are the benchmark results guaranteed for other workloads?
No. The figures are NVIDIA-reported outcomes from specified evaluations. Performance in other deployments may depend on the documents, languages, hardware, workload and integration choices, and independent testing was not included in the supplied source.
Primary source: Hugging Face · via ThorstenMeyerAI.com
