NVIDIA Nemotron 3 Embed Ranks #1 on RTEB Retrieval Benchmark

NVIDIA today released Nemotron 3 Embed, a collection of three open embedding models designed for production-scale retrieval, RAG, code retrieval, and agent memory. The flagship model, Nemotron-3-Embed-8B-BF16, ranks #1 on the RTEB leaderboard, scoring 78.5% on RTEB and 75.5% on MMTEB Retrieval. Two smaller 1B variants—Nemotron-3-Embed-1B-BF16 and Nemotron-3-Embed-1B-NVFP4—bring similar retrieval quality to cost- and latency-sensitive deployments.

The 1B BF16 model reduces error rate by 27% over its predecessor (llama-nemotron-embed-vl-1b-v2) on RTEB, scoring 72.4%, and by 28% on MMTEB Retrieval, scoring 71.0%. The NVFP4 variant, optimized for NVIDIA Blackwell hardware, delivers up to 2× higher throughput than BF16 while retaining 99%+ of BF16 retrieval accuracy with a smaller memory footprint. All three models feature a 32k context window for long-document retrieval, multilingual and code retrieval support, and come with open weights and fine-tuning recipes for domain adaptation.

The models were built by adapting the Ministral-3-8B-Instruct-2512 backbone, converting its causal decoder into a bidirectional encoder, and training on contrastive pre-training followed by fine-tuning on curated multilingual retrieval datasets across legal, finance, medical, business, and education domains. The 1B variants were created through a two-stage compression process: first, a 3B parent model derived from Ministral-3-3B-Instruct-2512 was compressed to 2B using NVIDIA ModelOpt's Neural Architecture Search engine, then to 1.14B through structured pruning and distillation from the 8B teacher. Final training employed a two-stage context-scaling schedule, first aligning multilingual behavior at 1024-token length, then expanding to 4096 tokens with long-context and reasoning datasets.

In agentic evaluations using Nemotron 3 Ultra as the search agent, stronger retrieval reduced estimated downstream agentic token cost per query across ViDoRe V3, BRIGHT, and BrowseComp-Plus. Better retrievers return relevant evidence earlier, enabling agents to complete tasks with fewer repeated searches and reasoning turns. The 8B model delivered both the highest average retrieval accuracy and the lowest estimated downstream token cost across these benchmarks. The models are available immediately on Hugging Face, deployable as NVIDIA NIM microservices, supported by vLLM, and accessible through leading AI Cloud and inference partners. Enterprise partners including Automation Anywhere, Boomi, IBM, Mem0, and Palantir are already evaluating the models for agentic retrieval, agent memory, code retrieval, and production inference workflows, with early results showing improvements over existing models.

Retrieval is foundational to agentic AI systems: when an agent cannot find relevant context quickly, it burns token budget on repeated searches and reasoning loops that carry noise forward. NVIDIA's Nemotron 3 Embed models address this by offering state-of-the-art retrieval accuracy across a range of deployment scenarios. The 8B flagship model's #1 RTEB ranking (78.5%) establishes the collection's quality ceiling, but the real production value lies in the 1B variants, which compress that quality into smaller footprints through structured pruning and knowledge distillation. The distillation strategy is noteworthy: the 1B models are not trained from scratch, but derived from a 3B Ministral backbone that undergoes two rounds of compression—first to 2B via NVIDIA's Neural Architecture Search engine, then to 1.14B via distillation from the 8B teacher. This pipeline allows the 1B BF16 to achieve 72.4% on RTEB (a 27% error reduction over its predecessor) while the NVFP4 variant adds Blackwell hardware acceleration for 2× throughput. The models carry production-ready features: 32k context windows for long documents and multi-turn agent histories, multilingual and code retrieval support, and fine-tuning recipes for domain adaptation. Early enterprise evaluations from Automation Anywhere, Boomi, IBM, Mem0, and Palantir suggest uptake momentum, with partners citing improvements in question-answering and agentic workflows.

No comments yet. Be the first to share your thoughts!