Nemotron 3 Embed brings retrieval models to agent memory

Nemotron 3 Embed is a new NVIDIA embedding family aimed at agentic retrieval: multilingual search, code retrieval, RAG, and agent memory. The release includes open weights, model cards, datasets, and training recipes, with an 8B BF16 model for quality testing and smaller 1B variants for deployment work. 1

The useful product split is simple: use Nemotron-3-Embed-8B-BF16 to test the retrieval ceiling, then check whether Nemotron-3-Embed-1B-NVFP4 holds up under your own latency, memory, and corpus constraints. The 8B model card reports a 32k max sequence length, 4096-dimensional vectors, and a 78.46 RTEB score; the 1B NVFP4 card describes 2048-dimensional vectors and query/document embedding for multilingual question answering over large text corpora. 2 3

First build step: treat it as an embedding swap. Index 100 representative docs with passage: embeddings, encode golden user questions with query: embeddings, then compare hit rate and latency against your current embedder before touching the rest of the stack. Benchmark rank is a starting signal, not a product KPI.