What changed
NVIDIA has announced the release of new open and commercially available models as part of its Nemotron Retriever collection. These models include Nemotron 3.5 Embed 8B BF16, 1B BF16, 1B NVFP4, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse at 631M parameters. These models are designed to enhance retrieval capabilities for enterprise agent workflows by supporting multimodal retrieval, relevance ranking, and document processing.
What this means in practice
The new models are intended to improve the efficiency of agents by reducing inference token costs and increasing retrieval accuracy. For example, 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. The models are also capable of retrieving evidence from both text and images, including tables, charts, infographics, datasheets, and schematics, allowing applications to search information conveyed by visual content alongside document text.
Limitations and availability
The models are available for deployment with different focus areas. Nemotron-3.5-Embed-8B-BF16 is suitable for quality-focused search over text and visually rich documents, while Nemotron-3.5-Embed-1B-BF16 is designed for multimodal indexing and query serving with a smaller model footprint. The 1B embedding model is available in BF16 and NVFP4, with quantization applied to both text and image encoders. The models are intended for use in enterprise settings and are part of the Nemotron Retriever collection. The supplied announcement does not specify exact deployment requirements or hardware compatibility beyond mentioning supported hardware for NVFP4 deployment.
