NVIDIA released the Nemotron 3 Embed model collection on July 16, 2026, targeting production-scale retrieval workflows such as RAG (Retrieval-Augmented Generation, fetching external data to improve AI responses) and agentic memory. The collection includes an 8B flagship model and two 1.14B variants optimized for varying deployment needs. The Nemotron-3-Embed-8B-BF16 model currently ranks #1 on the Retrieval-based Evaluation Benchmark (RTEB), scoring 78.5% and achieving 75.5% on the MMTEB retrieval benchmark.

For resource-constrained production environments, the collection offers the 1.14B BF16 model and the 1.14B NVFP4 variant. The NVFP4 model uses native 4-bit quantization for Blackwell GPU architectures, providing up to 2x higher throughput than the BF16 version while retaining 99% of its retrieval accuracy. These models support a 32k context window to facilitate retrieval across extensive technical documentation, multi-file code repositories, and long agent histories.

The models were developed by adapting the Ministral-3-8B-Instruct-2512 and 3B-Instruct-2512 backbones. The 1B variants were created through two rounds of structured pruning and distillation using NVIDIA's ModelOpt engine. NVIDIA has released open-source training recipes for fine-tuning and distillation, alongside availability on Hugging Face and as NVIDIA NIM (optimized inference microservices) containers. Various enterprise partners, including Automation Anywhere, Boomi, Mem0, and You.com, report using these models to improve semantic search and agentic reasoning efficiency.

NVIDIA released the Nemotron 3 Embed model collection on July 16, 2026, targeting production-scale retrieval workflows such as RAG (Retrieval-Augmented Generation, fetching external data to improve AI responses) and agentic memory. The collection includes an 8B flagship model and two 1.14B variants optimized for varying deployment needs. The Nemotron-3-Embed-8B-BF16 model currently ranks #1 on the Retrieval-based Evaluation Benchmark (RTEB), scoring 78.5% and achieving 75.5% on the MMTEB retrieval benchmark.

For resource-constrained production environments, the collection offers the 1.14B BF16 model and the 1.14B NVFP4 variant. The NVFP4 model uses native 4-bit quantization for Blackwell GPU architectures, providing up to 2x higher throughput than the BF16 version while retaining 99% of its retrieval accuracy. These models support a 32k context window to facilitate retrieval across extensive technical documentation, multi-file code repositories, and long agent histories.

The models were developed by adapting the Ministral-3-8B-Instruct-2512 and 3B-Instruct-2512 backbones. The 1B variants were created through two rounds of structured pruning and distillation using NVIDIA's ModelOpt engine. NVIDIA has released open-source training recipes for fine-tuning and distillation, alongside availability on Hugging Face and as NVIDIA NIM (optimized inference microservices) containers. Various enterprise partners, including Automation Anywhere, Boomi, Mem0, and You.com, report using these models to improve semantic search and agentic reasoning efficiency.