NVIDIA open-sources Nemotron 3 Embed models for AI agents and RAG

NVIDIA has open-sourced its new Nemotron 3 Embed model series, designed for AI agents and retrieval-augmented generation (RAG). The models are available with open weights for commercial use, free via Hugging Face and NVIDIA NIM.

These models have a 32K context window for high-precision retrieval, cutting down on repeated searches and token use. NVIDIA released three versions: an 8B-parameter BF16 model for high-accuracy enterprise tasks, a 1B-parameter BF16 model for low latency and cost, and a 1B-parameter NVFP4 model optimized for Blackwell GPUs, reducing memory use and boosting throughput.

In benchmarks, the 8B model scored 78.5% on RTEB and 75.5% on MMTEB retrieval. RTEB is a new standard from Hugging Face and the MTEB community, focused on evaluating real-world retrieval quality for RAG and agent scenarios.