NV-Embed

Overview

NV-Embed is a generalist embedding model that enhances the performance of decoder-only LLMs for embedding and retrieval tasks, with various architectural designs and training procedures.

Performance

Using only publicly available data, NV-Embed achieved:

• Record-high score of 69.32 on MTEB (Massive Text Embedding Benchmark)
• Ranked #1 on MTEB as of May 24, 2024
• Evaluated across 56 tasks

NVIDIA Llama-based Embedding Models (2026)

NVIDIA has developed several embedding models based on Llama architecture:

llama-text-embed-v2

• Built on Llama 3.2 1B architecture
• Optimized for high retrieval quality with low-latency inference

llama-3.2-nv-embedqa-1b-v2

• Dense text embedding model for fixed-length vector representations
• Note: API will be deprecated on 05/18/2026

llama-embed-nemotron-8b

• Open-weights text embedding model
• Achieves state-of-the-art performance on Multilingual MTEB leaderboard (October 21, 2025)
• Based on meta-llama/Llama-3.1-8B
• Fine-tuned version with bidirectional attention mechanism

NV-Embed-v2 (October 2025)

• Latest embedding model from NVIDIA
• Fine-tuned from Llama-3.1-8B
• Particularly powerful at understanding multilingual text
• Continues NVIDIA's leadership in embedding model performance

Key Innovations

• Improved techniques for training LLMs as generalist embedding models
• Decoder-only architecture optimization for embeddings
• Bidirectional attention mechanisms
• Multilingual capabilities

Applications

• Semantic search
• Retrieval-augmented generation (RAG)
• Multilingual information retrieval
• Cross-lingual tasks
• General-purpose text embedding

Availability

Available through:

• NVIDIA NIM (NVIDIA Inference Microservices)
• Hugging Face
• Research paper on arXiv