NVIDIA's open-source Nemotron 3 Embedded text embedding model series

NVIDIA's open-source Nemotron 3 Embed series of text embedding models is designed for retrieval enhancement and agent retrieval. The model is available in 8B and 1B parameter scales, supporting 32k context windows and 34 languages. The flagship 8B model achieved a score of 78.5% on the RTEB multilingual retrieval benchmark, ranking first. The 1B version, through pruning and distillation, significantly reduces inference costs while retaining 95% retrieval accuracy.