Nvidia Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval (huggingface.co)

Nvidia has launched the Nemotron 3 Embed, a set of open embedding models that significantly enhances retrieval capabilities, achieving top rankings on the Retrieval Evaluation Benchmark (RTEB). This collection includes three models: the flagship 8B variant, which excels in precision-critical retrieval, and two 1B variants designed for high-efficiency and hardware-accelerated deployments. The models are particularly relevant for production-scale applications such as retrieval-augmented generation (RAG), agentic retrieval, and code retrieval. They boast improvements in tuning and fine-tuning options, supporting multilingual data and extensive context handling, with some models offering a 32k context window to reduce truncation issues. This release is noteworthy for the AI/ML community because it allows developers greater flexibility and efficiency in deploying state-of-the-art retrieval models across diverse operational environments. The 8B model achieves a high retrieval accuracy, scoring 78.5% on the RTEB, while the 1B models retain significant performance gains even at lower costs and faster deployment speeds. Moreover, the integration of features like open weights, fine-tuning recipes, and a microservice architecture ensures that organizations can easily adapt these models to their infrastructure needs, paving the way for enhanced AI interactions and efficiency in information retrieval tasks across various industries.