NVIDIA Expands Nemotron 3.5 with New Retrieval Models
NVIDIA has announced the expansion of its Nemotron collection, introducing open, commercially available models designed to advance enterprise agent workflows through multimodal retrieval, relevance ranking, and document processing.
Radwan AbdulhadiEditor and technology writer
Expanding the Nemotron Retriever Collection
An AI agent’s answers depend directly on the evidence it can retrieve from its environment. When evidence is incomplete or irrelevant, agents often repeat searches and inspect extra documents, consuming valuable inference tokens before completing a given task. To address these enterprise-scale inefficiencies, NVIDIA has expanded its <a href="https://huggingface.co/blog/nvidia/nemotron-retriever-world-class-agentic-retrieval">Nemotron Retriever</a> collection with new open and commercially available models tailored for embedding, reranking, and document extraction.
The newly introduced models encompass the Nemotron 3.5 Embed 8B BF16, 1B BF16, and 1B NVFP4 variations, alongside the Nemotron 3.5 Rerank 8B BF16 model and Nemotron 3.5 Parse at 631M parameters. Together, these systems aim to bring advanced multimodal retrieval, relevance ranking, and document processing directly to enterprise agent workflows, helping agents reach grounded answers with fewer wasted tokens.
Multimodal Retrieval and Leaderboard Performance
Retrieval quality significantly affects both the answers an agent produces and its overall large language model inference cost. Relevant evidence helps the agent decide its next steps more effectively, while reranking ensures higher-quality evidence enters the context window. The <a href="https://huggingface.co/nvidia">nvidia</a> team designed these models to process text alongside document images—such as tables, charts, infographics, datasheets, and schematics—enabling comprehensive information discovery across diverse visual and textual data formats.
On the ViDoRe v3.1 leaderboard, the Nemotron 3.5 Embed 8B BF16 model achieved a score of 61.17 nDCG@10, ranking first among dense retrieval models. This marks an 8.94-point improvement over previous model iterations. Meanwhile, the distilled 1B BF16 model scored 58.38 nDCG@10, retaining approximately 95.4% of its larger teacher model's score while offering a more compact deployment option.
Evaluation Across Benchmarks and Languages
To evaluate how retrieval configuration impacts an agent's ability to combine evidence and complete complex tasks, NVIDIA tested Nemotron 3.5 Embed and Rerank across multiple evaluation datasets, including ViDoRe v3, BRIGHT, and BrowseComp-Plus. The experiments measured task success rates alongside total agent inference tokens per task across various search tool implementations.
The evaluation results demonstrated that adding Nemotron 3.5 Rerank reduces agentic token costs per query by approximately 14% when paired with Nemotron 3.5 Embed 8B, and by 13% with Nemotron 3.5 Embed 1B, while concurrently improving retrieval accuracy. Furthermore, multilingual evaluations showed strong performance, with the 8B model achieving 72.6 nDCG@10 on MIRACL Vision and outperforming competitive baselines across multiple language settings.
Refining Context with Nemotron 3.5 Rerank
Embedding models make large document collections searchable by representing content as vectors. Once initial retrieval returns a candidate set, the Nemotron 3.5 Rerank 8B BF16 model evaluates each candidate against the query to refine the ordering across both text and image inputs. By distinguishing broadly related content from precise evidence, the reranker helps agents optimize their context budgets.
In collaboration with the MTEB team, NVIDIA also launched the ViDoRe v3.1 reranking leaderboard to benchmark rerankers on document retrieval. On this leaderboard, Nemotron 3.5 Rerank 8B BF16 achieved a score of 60.49 nDCG@10 using combined image-and-text inputs, establishing a strong benchmark for agentic retrieval systems seeking to surface the strongest evidence early in the reasoning process.
Sources
- Hugging Face BlogNVIDIA Nemotron 3.5 Embed, Rerank, and Parse: Advancing Agents with World-Class Retrieval Models
Continue chronologically
Related entity coverage
Related stories
NVIDIA Fine-Tunes Nemotron for Gold-Level IMO and IOI Results
NVIDIA demonstrates how fine-tuning the Nemotron model family using SFT, RL, and inference loops achieved gold-medal levels at IMO 2026 and IOI 2026.
NVIDIA and AI4Bharat Launch Indic Pretraining Corpus
NVIDIA and IIT Madras lab AI4Bharat introduce Nemotron-pretraining-Indic-v1, featuring 2.8 trillion tokens and an open data generation pipeline.
NVIDIA Launches Nemotron 3.5 Transcribe and Granary v2
NVIDIA introduces Nemotron 3.5 Transcribe and Granary v2, delivering advanced multilingual transcription, speaker attribution, and promptable audio tools.
NVIDIA Launches Nemotron 3 Diarization Open-Weight Model
NVIDIA introduces Nemotron 3 Diarization, a 100M-parameter open-weight model topping VoiceArena's leaderboard for real-time multi-speaker AI.
