What was released
NVIDIA released the Nemotron 3 Embed family on July 16, 2026: an accuracy-oriented 8B BF16 model, an efficient 1B BF16 model, and a 1B NVFP4 variant for Blackwell throughput. It also provides open weights, datasets, and fine-tuning and distillation recipes.
Why retrieval agents care
Embedding models are the entry point for RAG and agent memory, locating relevant evidence in documents, code, and conversation history. NVIDIA reports multilingual and code retrieval, a 32K context window, and a score of 78.5 for the 8B model at the top of the July 15 RTEB snapshot.
Three deployment profiles
- 8B BF16 for precision-critical retrieval and enterprise RAG
- 1B BF16 for lower-latency, cost-sensitive serving
- 1B NVFP4 for Blackwell throughput and reduced memory use
Limits
The rankings and efficiency figures come from NVIDIA's evaluation and a leaderboard snapshot. Teams should test their own languages, domains, freshness, recall, downstream answer quality, indexing cost, and end-to-end latency. Model context claims and the validated limit of a particular NIM runtime should also be checked by version.