source: hugging face blog: nvidia nemotron 3 embed ranks #1 overall on rteb, advancing agentic retrieval
level: technical
nvidia released nemotron 3 embed, a set of open embedding models for retrieval tasks. the collection includes an 8b flagship model and two 1b variants. the 8b model ranks first on the rteb leaderboard and scores 78.5% on rteb and 75.5% on mmteb retrieval. the 1b bf16 model scores 72.4% on rteb, cutting error rate by 27% over its predecessor. a blackwell-optimized nvfp4 variant offers high throughput with a smaller memory footprint.
the models support a 32k context window, multilingual and code retrieval, and fine-tuning recipes. they are available on hugging face, as nvidia nim microservices, and through cloud partners. evaluations show stronger retrieval reduces downstream token costs in agentic workflows. the 8b model achieved the highest accuracy and lowest estimated token cost across vidore v3, bright, and browsecomp-plus benchmarks. the nvfp4 variant retains over 99% of bf16 accuracy while doubling throughput on blackwell gpus.
the 8b model adapts a ministral-3-8b backbone into a bidirectional encoder, trained with contrastive pre-training and fine-tuned on multilingual data. the 1b model was created by pruning a 3b retriever to 2b then 1.14b using structured pruning and distillation from an 8b teacher. training used a two-stage context scaling schedule up to 4096 tokens. enterprise partners like automation anywhere, boomi, and palantir report strong early results in agentic retrieval and memory tasks.
why it matters: better retrieval models can lower token costs and improve accuracy in ai agent workflows, making them more practical for enterprise use.
source: hugging face blog: nvidia nemotron 3 embed ranks #1 overall on rteb, advancing agentic retrieval