NVIDIA launches Nemotron-3-Embed retrieval models

NVIDIA has released the Nemotron-3-Embed family, a pair of open, commercially usable multilingual embedding models aimed at agentic retrieval, code retrieval, and agent memory. The launch matters because NVIDIA is positioning retrieval quality as a direct lever on agent answer quality, and because its larger model was announced as the top overall system on RTEB.

What was released

According to NVIDIA and follow-up posts, the series includes two BF16 models: Nemotron-3-Embed-1B-BF16 and Nemotron-3-Embed-8B-BF16. The models target sentence similarity and text embedding use cases, with a multilingual focus. @tomaarsen also noted that the models are natively integrated into Sentence Transformers, where users can directly load `nvidia/Nemotron-3-Embed-8B-BF16` for query and document encoding.

Benchmark results and product positioning

NVIDIA said Nemotron-3-Embed-8B-BF16 ranked overall No.1 on RTEB, a benchmark the company described as measuring retrieval accuracy on real-world tasks. A repost by @kimmonismus cited the reported scores as 78.46 average NDCG@10 on RTEB and 75.45 on MMTEB Retrieval. NVIDIA’s framing is that stronger retrieval gives agents more relevant context, which should improve downstream answer accuracy; posts around the release also described the model as well suited to cost-conscious agentic retrieval. On July 18, @nvidia added that the 8B-BF16 model had entered Hugging Face Trending.

Compression and deployment notes

Beyond the headline 8B result, @tomaarsen highlighted an unusual compression pipeline for Nemotron-3-Embed-1B: starting from a 3B parent model, NVIDIA ModelOpt was used for NAS pruning down to 2B, followed by distillation from an 8B teacher using cosine and MSE losses. Separately, in a repost, @soumitrashukla9 focused on the 1B NVFP4 variant, saying the post’s description suggested its footprint could remain manageable in many scenarios even when KV and CUDA overhead were included.

2026-07-17 ~ 2026-07-18 · 11 related posts

2 near-duplicate retellings: NVIDIAAI · NVIDIAAI