NVIDIA Releases Nemotron 3 Embed: Top-Ranked Open Embedding Models

NVIDIA has introduced Nemotron 3 Embed, a new collection of open-source embedding models designed to enhance production-scale retrieval-augmented generation (RAG), agentic retrieval, and code retrieval tasks. The collection features three distinct checkpoints, including an 8B parameter model that has achieved the number one ranking on the Retrieval Embedding Benchmark (RTEB) with a score of 78.46.

Model Architecture and Capabilities

The Nemotron 3 Embed collection consists of three transformer encoders trained with bidirectional attention masking. The 8B variant, Nemotron-3-Embed-8B-BF16, is built upon the Ministral-3-8B-Instruct-2512 base. For smaller footprint requirements, NVIDIA provides two 1.14B parameter variants based on the Ministral-3-3B-Instruct-2512 model. All three checkpoints support a maximum sequence length of 32,768 tokens, allowing agents to process extensive conversation histories and long-form documents without the need for aggressive chunking.
The models are released under the OpenMDW-1.1 license and were evaluated across 34 languages. The final embeddings are generated using average pooling over token-level representations, ensuring high-quality semantic retrieval across diverse applications, including multilingual enterprise search and natural-language-to-code lookup.

Optimization and Performance

NVIDIA achieved the 1B model’s performance through a rigorous compression pipeline rather than a standard training run. The process utilized NVIDIA ModelOpt’s Neural Architecture Search to prune the 3B parent model, followed by iterative rounds of distillation using cosine distance and mean squared error loss. This approach resulted in a 10.4-point gain on the RTEB benchmark compared to the prior-generation baseline.
The collection also includes the Nemotron-3-Embed-1B-NVFP4 checkpoint, which is optimized for the Blackwell architecture. By utilizing 4-bit quantization, this version delivers up to two times higher throughput than the BF16 variant while retaining over 99% of its retrieval accuracy. The NVFP4 model also supports dynamic embedding sizes, allowing developers to slice the 2048-dimensional vectors to 1024 or 512 dimensions as needed for specific deployment requirements.

Deployment and Integration

To support developers, NVIDIA provides multiple integration paths, including compatibility with Transformers and Sentence Transformers libraries. For high-performance serving, the models can be deployed via vLLM, and NVIDIA has released an optimized NIM microservice for the 1B model that demonstrates superior performance on GB200 and RTX PRO 6000 hardware.
The models utilize specific input prefixes—query: for search queries and passage: for documents—to ensure accurate retrieval. Because the embeddings are L2-normalized, the dot product between query and document vectors functions as a direct measure of cosine similarity. This streamlined approach allows for flexible implementation, such as cost-tiered RAG architectures where high-volume recall is handled by the 1B-NVFP4 model, while more complex queries are routed to the 8B-BF16 checkpoint.