Nvidia has released Nemotron 3 Embed, a family of open-source embedding models that immediately claim the top spot on the retrieval benchmark leaderboard.
Released on July 17 under the permissive OpenMDW-1.1 license, the flagship 8B BF16 model scores 78.5% on the Retrieval Text Encoding Benchmark (RTEB) and 75.5% on the Massive Multilingual Text Embedding Benchmark (MMTEB) Retrieval — the highest scores recorded on both.
What Makes It Different
Embedding models convert text into numerical vectors that capture semantic meaning, powering the retrieval half of retrieval-augmented generation (RAG) — the technique that lets AI systems search through documents, databases and knowledge bases before answering.
Nemotron 3 Embed supports a 32,000-token context window across 34 languages, making it one of the most capable multilingual embedding models available. All three checkpoints — 8B BF16, 8B FP8, and a 1B NVFP4 variant optimized for Nvidia's next-generation Blackwell architecture — are available for download.
Blackwell Optimization
The 1B NVFP4 variant is particularly notable: it retains 99%+ of the BF16 model's accuracy while running at roughly 2x the throughput on Blackwell hardware, demonstrating Nvidia's push to optimize its software stack for its own next-generation silicon before that hardware even ships in volume.
The Embedding Model Race
Nemotron 3 Embed enters a rapidly intensifying market for embedding models. As enterprises increasingly adopt RAG architectures to ground AI systems in their own data, the quality of the embedding model directly determines whether a RAG pipeline finds relevant information or returns noise.
Nvidia's decision to release the model under a permissive open license positions it as a direct competitor to OpenAI's text-embedding-3, Cohere's Embed v3, and open alternatives like BGE and E5.
The models are available on Hugging Face and through Nvidia's NGC catalog, with integrations for major vector databases including Pinecone, Weaviate, Qdrant and Milvus.
