NVIDIA has released Nemotron 3 Embed, a collection of three open embedding models (8B and 1B variants) that top the RTEB multilingual retrieval leaderboard. The flagship 8B model scores 78.5% on RTEB, while the 1B variants offer efficient deployment options. Key features include a 32k context window, multilingual and code retrieval support, NVFP4 quantization for NVIDIA Blackwell GPUs delivering 2x throughput over BF16, and open-source fine-tuning and distillation recipes via NeMo AutoModel. The models are built by converting causal decoders into bidirectional encoders and compressing via structured pruning and knowledge distillation. Evaluations show that better retrieval directly reduces downstream agentic token costs. Enterprise partners including IBM, Zoom, Palantir, and ServiceNow are already evaluating the models for production RAG and agentic retrieval workloads.
