Source Posts on X
NVIDIA’s new #1 RTEB embedding model is built around a simple idea: the cheapest reasoning token is the one an agent never generates. Nemotron-3-Embed-8B reaches 78.46 average NDCG@10 on RTEB and 75.45 on MMTEB Retrieval. NVIDIA tested it inside a search agent powered by Nemotron 3 Ultra. More accurate retrieval returned relevant evidence earlier, reducing repeated searches, context inspection and reasoning turns. The result: NVIDIA says the 8B model achieved the highest retrieval accuracy and lowest estimated downstream token cost across three agentic evaluations. For cheaper deployment, NVIDIA also released two 1B models. The Blackwell-optimized NVFP4 version offers up to 2× BF16 throughput while retaining more than 99% of its retrieval quality. All three support 32K context, multilingual and code retrieval, enterprise RAG and agent memory. The weights, data and recipes are open for commercial use. Agent cost optimization is moving into the retrieval layer.