NVIDIA's Nemotron 3 Embed 8B tops RTEB retrieval leaderboard

Embedding models rarely get the release-day attention that chat models do, and that is usually the mistake. Retrieval quality is what decides whether a RAG system answers the right question or a plausible-sounding wrong one, and this week NVIDIA dropped three open embedding checkpoints that reportedly sit at the top of the current retrieval benchmark.

According to MarkTechPost's writeup, the Nemotron 3 Embed collection ships three sizes: an 8B BF16 flagship, a 1B BF16, and a 1B NVFP4 quantized variant. The 8B checkpoint posts 78.46 average NDCG@10 on RTEB, described as a #1 overall ranking, and 75.45 on MMTEB Retrieval. Every checkpoint takes up to 32,768 tokens of context and was evaluated across 34 languages. The 1B NVFP4 variant is the interesting hardware story: the claim is up to 2x higher throughput than BF16 while retaining 99%+ of the BF16 retrieval accuracy, with support spanning Ampere, Hopper, Lovelace, and Blackwell.

The commercial angle sits in the license. The collection ships under OpenMDW 1.1, and a top-of-leaderboard retrieval model that customers can self-host on hardware they already own changes the math against the paid embedding APIs, especially for teams whose data cannot leave their own environment. Retrieval is the unglamorous half of RAG, and moving it in-house is often where the operating cost actually lives.

The honest caveat is that RTEB and MMTEB averages do not tell you how the model behaves on your corpus, your query mix, or your domain vocabulary. The 34-language evaluation is broad but the reporting does not break out how low-resource languages fare against English, and the fine print on OpenMDW 1.1 is worth reading before anyone commits it to a production stack. What the writeup does not give you is latency at realistic batch sizes, or how these checkpoints hold up on niche domains like legal, medical, or code retrieval.

The teams best positioned to move on this quickly are the ones that already run NVIDIA fleets and were paying per token for an embedding API they did not love. A Blackwell-optimized 1B in NVFP4 is a very different unit economic than an external call, and that is the shift worth pressure-testing on your own data before the next vendor renewal.