NVIDIA released Nemotron 3 Embed, a family of open embedding models that achieved state-of-the-art retrieval across the accuracy-efficiency curve, led by an 8B model that ranks #1 on the RTEB leaderboard. The collection also includes efficient 1B BF16 and NVFP4 variants designed for production-scale RAG, agentic retrieval, code retrieval, and agent memory.

The 8B flagship scores 78.5% on RTEB and 75.5% on MMTEB Retrieval. The 1B BF16 model scores 72.4% on RTEB (reducing error rate by 27% over its predecessor) and 71.0% on MMTEB Retrieval (reducing error rate by 28%). All models support a 32k context window, multilingual and code retrieval, and are available with open weights, datasets, and training recipes.

For agentic retrieval evaluation, the authors built a search agent powered by Nemotron 3 Ultra and varied the embedding model. Stronger retrieval reduced downstream agentic token cost by returning relevant evidence earlier, reducing repeated searches and reasoning turns. The 8B model delivered both the highest average retrieval accuracy and the lowest estimated downstream token cost across ViDoRe V3, BRIGHT, and BrowseComp-Plus.

The 1B NVFP4 variant on Blackwell delivers up to 2x higher throughput than BF16 while retaining 99%+ retrieval accuracy. Optimized NIM microservices for the 1B model match or outperform vLLM checkpoints on GB200 and RTX PRO 6000 GPUs.

The 8B model adapts a Ministral-3-8B-Instruct backbone by converting its causal decoder into a bidirectional encoder, trained with contrastive pre-training on web and synthetic text pairs, then fine-tuned on curated multilingual datasets. The 1B model was derived through structured pruning (using ModelOpt’s NAS engine) and distillation from 8B teacher checkpoints in a two-round compression pipeline, followed by progressive two-stage context scaling (1024 then 4096 tokens).

Enterprise partners including Automation Anywhere, Boomi, IBM, Mem0, Palantir, ServiceNow, turbopuffer, You.com, Zep, and Zoom provided positive early evaluations. The models are available immediately on Hugging Face, as NVIDIA NIM microservices, through vLLM, and via AI Cloud partners including Baseten, Bitdeer AI, DeepInfra, Friendli AI, and OpenRouter. NeMo AutoModel fine-tuning and distillation recipes are also open-sourced; on the NV Docs evaluation, fine-tuning the 1B BF16 model improved NDCG@10 from 56.7% to 63.3% (+11.6%).