NVIDIA Open-Sources Nemotron 3 Embed: 8B Tops RTEB, NVFP4 Nearly Doubles Throughput on Blackwell at Full Accuracy
NVIDIA released the open Nemotron 3 Embed collection; the 8B ranks #1 on RTEB at 78.46 average NDCG@10, while a distilled 1B and an NVFP4 build retain 99%+ retrieval accuracy at up to 2x throughput.
An open embedding collection that tops RTEB
NVIDIA released Nemotron 3 Embed on July 15–16, 2026, with three open checkpoints: Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, and Nemotron-3-Embed-1B-NVFP4. The 8B checkpoint ranks #1 on the RTEB retrieval benchmark at 78.46 average NDCG@10.
How the small model was made: NAS pruning + distillation
The 1B version was not trained from scratch; it came from the 8B teacher via ModelOpt NAS (neural architecture search) pruning plus COS+MSE distillation — inheriting as much of the 8B's retrieval quality as possible in a smaller footprint.
NVFP4: near-lossless quantized speedup
The NVFP4 build, targeting NVIDIA's Blackwell architecture, retains 99%+ of BF16 retrieval accuracy while delivering up to 2x throughput. On the same hardware, that meaningfully cuts vectorization cost and latency for retrieval/RAG workloads with almost negligible accuracy loss.
Specs and license
All three checkpoints support 32,768-token inputs, fitting long-document retrieval, under the OpenMDW-1.1 license. For applications built on RAG and semantic search, a collection that tops the benchmark while also offering a 1B small model and an NVFP4 quantization lets teams trade off accuracy, size and throughput as needed — lowering the bar to running a self-hosted embedding service.