Nemotron 3 Embed - NVIDIA's Open-Source Text Embedding Model Series

What is Nemotron 3 Embed

Nemotron 3 Embed is a series of open-source text embedding models from NVIDIA, designed for retrieval-augmented generation and agentic retrieval. The models come in two parameter scales, 8B and 1B, support a 32k context window and 34 languages, and are released under the OpenMDW-1.1 license for commercial use. The flagship 8B model ranks first on the RTEB multilingual retrieval leaderboard with a score of 78.5%, while the 1B version, through pruning and distillation, retains 95% of the retrieval accuracy while greatly reducing inference costs.

Main Features of Nemotron 3 Embed

• Multilingual dense retrieval: Generates text embeddings covering 34 languages and supports cross-lingual semantic retrieval.
• Long-context processing: A 32k context window supports long-document retrieval and question answering.
• Flexible dual-spec deployment: The 8B version pursues ultimate retrieval accuracy, while the 1B version balances cost and speed.
• Hardware-optimized quantization: The NVFP4 quantized version is optimized for the NVIDIA Blackwell architecture, increasing throughput by up to two times.
• Production-grade deployment support: Provides NVIDIA NIM microservices, vLLM, and native Hugging Face integration.

Technical Principles of Nemotron 3 Embed

• Based on Ministral-3-8B-Instruct-2512: The 8B model is fine-tuned and optimized for retrieval tasks on top of an open-source architecture.
• Contrastive learning training: Maps text into a dense vector space through large-scale contrastive learning, bringing semantically similar text closer together in the vector space.
• Structured pruning and distillation: The 1B model removes redundant parameters and distills the behavior of the 8B model, compressing to 1B while retaining 95% retrieval accuracy.
• NVFP4 low-precision quantization: A 4-bit floating-point quantization technique for the Blackwell architecture that achieves a two-times throughput increase with almost no accuracy loss.

Follow WeChat and reply “Open Source” to join theAI Open Source Project Exchange Group

How to Use Nemotron 3 Embed

• Hugging Face download: Visit the Hugging Face model repository to download the weights directly, and use the Transformers or Sentence-Transformers library to load and run inference.
• NVIDIA NIM microservices: Obtain NIM microservice containers through build.nvidia.com to quickly deploy production-grade inference services in enterprise environments.
• vLLM deployment: Deploy the model with vLLM's high-throughput inference engine to support large-scale concurrent retrieval requests.
• Fine-tuning and distillation: Refer to the fine-tuning recipes provided by NVIDIA to adapt the model to your own domain data; or use the distillation recipe to transfer knowledge from the 8B model to a smaller model.

Core Advantages of Nemotron 3 Embed

• Ranked first on RTEB: The 8B model surpasses all open-source and closed-source models on the RTEB multilingual retrieval benchmark with a score of 78.5%.
• Open source and commercially usable: The model uses the OpenMDW-1.1 license, with weights, fine-tuning, and distillation recipes fully open, supporting commercial scenarios.
• Hardware co-optimization: The NVFP4 quantized version is deeply adapted to the Blackwell architecture, achieving the best balance between accuracy and throughput.
• Long-context support: The 32k context window is leading among embedding models, meeting the needs of long-document retrieval.

Project address of Nemotron 3 Embed

• Project official website：https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb
• HuggingFace model repository：https://huggingface.co/collections/nvidia/nemotron-3-embed

Comparison of Nemotron 3 Embed with similar competing products

Application scenarios of Nemotron 3 Embed

• Enterprise knowledge base Q&A: Build a multilingual RAG system based on internal documents, supporting employees in cross-language retrieval of technical documents and rules and regulations.
• Intelligent customer service retrieval: Provide customer service robots with high-precision semantic retrieval capabilities to quickly match user questions with historical solutions.
• Legal and financial compliance retrieval: Locate key clauses in lengthy contracts, financial reports, and regulatory documents to assist compliance review and risk analysis.
• E-commerce product semantic search: Understand users' search intent through semantics to achieve cross-language product matching and recommendation.
• Code and documentation retrieval: Retrieve related functions, API descriptions, and implementation examples in large codebases and technical documentation.