NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB

Embedding models determine which passages an AI agent retrieves from data, and NVIDIA released Nemotron 3 Embed, a collection of open embedding models designed for production-scale retrieval tasks across multiple languages. The largest model in the collection ranks first on RTEB, the Retrieval Embedding Benchmark, and comes in three variants ranging from 1 billion to 8 billion parameters with different optimization approaches. The smaller models were created through compression techniques including pruning and distillation rather than training from scratch, achieving high accuracy retention while reducing computational costs. These models support applications including multilingual search, code retrieval, long-context agent memory, and cost-tiered systems that route queries between smaller and larger models based on difficulty.