Eleven Realtime Translation
Translate live speech into spoken audio in more than 90 languages through an enterprise WebSocket API.
NVIDIA Nemotron 3.5 Embed 8B
huggingface.co
Create multilingual embeddings for semantic search across document text, page images and combined image-text content.
Visit website
Pricing
Free
Available on
Linux, NVIDIA CUDA, NVIDIA NIM API
Category
Developer Tools
What is NVIDIA Nemotron 3.5 Embed 8B?
NVIDIA Nemotron 3.5 Embed 8B is an embedding model for multilingual semantic search and document retrieval. It accepts document text, page images or combined image-text inputs, producing vectors that can be matched against text queries.
The downloadable checkpoint is Nemotron-3.5-Embed-8B-BF16. Search engineers can obtain its weights from Hugging Face or access the model through NVIDIA NIM to build retrieval components for search and RAG applications.
Key features
Visual document retrieval
Embed pages containing text, tables, charts and infographics, including combined image-text documents.
Multilingual retrieval
Text retrieval was evaluated across 34 languages and image retrieval across 18 languages. These counts describe evaluation coverage.
Adjustable vector dimensions
Generate normalized 4,096-dimensional vectors, or truncate and renormalize them for shorter representations such as 1,024 or 512 dimensions.
Local deployment options
The checkpoint includes examples for Transformers, Sentence Transformers and vLLM.
NIM access
NVIDIA provides a model endpoint and API reference for developers evaluating hosted embedding workflows.
Who is it for?
Search across visual reports
Search engineers could retrieve relevant pages from reports containing charts, tables and prose.
Multilingual knowledge retrieval
Enterprise developers could build semantic search over document collections in multiple languages.
Retrieval for RAG applications
Application teams could supply relevant document passages or page images to a separate answer-generation system.
How to use NVIDIA Nemotron 3.5 Embed 8B
- Choose the downloadable BF16 checkpoint or NVIDIA's hosted endpoint.
- For local deployment, follow a supported runtime example and prepare a compatible CUDA environment.
- Embed document text or page images and store the vectors with their original content in a vector database.
- Embed text queries and retrieve documents using vector similarity.
NVIDIA Nemotron 3.5 Embed 8B pricing
Free
Model materials are available free of charge under OpenMDW-1.1, subject to its terms. NVIDIA also lists a free endpoint for evaluation. Compute, hosting and production services are separate considerations.
Check current pricing ↗What to know before you use it
- Queries accept text only. The model returns embeddings rather than parsed documents or generated answers.
- Context is limited to 32,768 tokens. Image longest-edge resolution defaults to 1,120 pixels and can increase to 1,540 pixels.
- The documented BF16 examples require a CUDA GPU; this is infrastructure for developers rather than a standalone search application.
- The hosted API is a preview trial service governed by NVIDIA API Trial Terms, separately from the downloadable model license.
Frequently asked questions
Can I use the model commercially?
Yes. NVIDIA permits commercial and non-commercial use under OpenMDW-1.1. Redistribution requires retaining the agreement and applicable notices.
Can it search charts and tables?
Yes. Document-page inputs can contain charts, tables, infographics and text, retrieved using text queries.
What else does a complete retrieval application need?
A typical system stores document embeddings in a vector database linked to the original content. RAG applications also need a separate model to generate answers.