NVIDIA has introduced NV-Embed, a generalist embedding model that has achieved the top position on the Massive Text Embedding Benchmark (MTEB) leaderboard. NV-Embed, which utilizes decoder-only large language models (LLMs), has outperformed BERT and T5-based models in general-purpose text embedding tasks. The model incorporates several architectural improvements, including latent attention pooling and a two-stage contrastive instruction-tuning. NV-Embed scored 59.36 on 15 retrieval tasks within the MTEB benchmark, using only publicly available data. This development is expected to enhance the performance of decoder-only LLMs like Mistral-7B, maintaining simplicity and reproducibility. The model was developed by a team including C Lee, R Roy, M Xu, and J Raiman.