NVIDIA Nemotron 3 Embed 8B Tops RTEB Leaderboard with 78.5% Score, 1B Variant Cuts Error Rate 27%

NVIDIA Nemotron 3 Embed 8B Tops RTEB Leaderboard

NVIDIA's Nemotron-3-Embed-8B-BF16 ranks #1 on the RTEB leaderboard with a 78.5% score, according to the company's announcement on July 16, 2026. The release includes three open-weight embedding models designed for production RAG, agentic retrieval, and code retrieval systems.

Model Lineup and Benchmark Results

The collection includes:

• Nemotron-3-Embed-8B-BF16: Flagship model scoring 78.5% on RTEB and 75.5% on MMTEB Retrieval
• Nemotron-3-Embed-1B-BF16: Scores 72.4% on RTEB and 71.0% on MMTEB Retrieval, reducing error rate by 27% over its 1B predecessor (llama-nemotron-embed-vl-1b-v2)
• Nemotron-3-Embed-1B-NVFP4: Blackwell-optimized 4-bit variant retaining 99%+ of BF16 accuracy while delivering up to 2x higher throughput

NVIDIA evaluated the models across multiple benchmarks using average NDCG@10 scores, including ViDoRe V3 Text, MMTEB Retrieval, and LongEmbed.

Technical Specifications

All three models feature:

• 32,000 token context window
• Multilingual and code retrieval capabilities
• Open weights with training recipes
• Day-0 integration with Hugging Face, vLLM, and NVIDIA NIM microservices

The 8B model adapts the Ministral-3-8B-Instruct-2512 backbone by converting its causal decoder into a bidirectional encoder. According to NVIDIA, the model underwent contrastive pre-training on web-sourced and synthetic text pairs, followed by fine-tuning on curated multilingual retrieval datasets across legal, finance, medical, business, and education domains.

Agentic Retrieval Performance

NVIDIA tested the models in an agentic workflow using a search agent powered by Nemotron 3 Ultra. The evaluation measured retrieval accuracy against estimated downstream token cost across ViDoRe V3, BRIGHT, and BrowseComp-Plus benchmarks. The company claims that stronger retrieval reduces token cost by returning relevant evidence earlier, helping agents avoid repeated searches and extra reasoning turns.

Production Deployment Options

The NVFP4 variant targets high-throughput deployments on NVIDIA Blackwell architectures. NVIDIA reports that the optimized NIM microservice for the 1B model matches or outperforms vLLM on GB200 and RTX PRO 6000 GPUs across input sequence lengths of 256 and 1,024 tokens.

Pricing for the models was not disclosed. The models are available immediately through Hugging Face and NVIDIA's deployment platforms.

What This Means

NVIDIA's RTEB #1 ranking establishes a new benchmark target for open embedding models, though real-world performance will depend on domain-specific use cases. The 27% error reduction in the 1B model addresses a practical deployment constraint: teams often sacrifice retrieval quality for latency and cost, but this gap is narrowing. The Blackwell-optimized NVFP4 variant represents a hardware-software co-design approach that may become standard for production embedding deployments, assuming Blackwell adoption accelerates in enterprise infrastructure.