NVIDIA released Nemotron 3 Embed, a family of open embedding models built for enterprise retrieval and agentic AI systems, with the flagship 8B model topping the RTEB leaderboard at a 78.5% score. Companion 1B variants target efficient production deployment, including an NVFP4 version that delivers double the throughput on NVIDIA Blackwell hardware. The models support a 32k context window and multiple languages, and were trained via bidirectional adaptation of Ministral backbones combined with contrastive pre-training and distillation.
