NVIDIA released Nemotron 3 Embed on July 16, and the 8B model immediately landed at #1 on RTEB — the Retrieval Embedding Benchmark that tests real-world search tasks across languages. That would be noteworthy on its own. What makes it significant is everything that comes with it: Apache 2.0 license, three deployment-ready checkpoints, and a published fine-tuning recipe. Proprietary embedding APIs just lost their main differentiator.
What RTEB Actually Measures #
RTEB is not another synthetic benchmark. It tests retrieval accuracy across real tasks — document search, agent memory, code retrieval, multilingual queries — the kind of work that actually happens in production. Nemotron-3-Embed-8B scores 78.5% on RTEB and 75.5% on MMTEB Retrieval, topping Voyage 4 Large, OpenAI text-embedding-3-large, and Cohere embed-v4.
One caveat worth noting: Google Gemini Embedding 001 still leads the English-only MTEB leaderboard at 68.32%. If your use case is English-only and you are already embedded in Google’s stack, this is not an automatic swap. But for multilingual workloads, agentic retrieval, and agent memory, Nemotron is the new benchmark to beat.
Three Models, One Decision #
NVIDIA ships three checkpoints, and picking the right one takes about thirty seconds:
Nemotron-3-Embed-8B-BF16— Maximum retrieval accuracy, 4,096-dimension embeddings, 32,768-token context window. Use this when quality is the only variable that matters.Nemotron-3-Embed-1B-BF16— 95% of the 8B model’s quality at a fraction of the inference cost. This is the default starting point for most teams. It scores 72.4% on RTEB, which still beats every major proprietary model.Nemotron-3-Embed-1B-NVFP4— Same 1B model, quantized for Blackwell GPUs (H100, GB200). Retains 99%+ of BF16 accuracy with up to 2x throughput. If you are on Blackwell hardware, there is no reason to run BF16.
All three share the 32,768-token context window — a notable edge over OpenAI text-embedding-3’s 8,192-token limit. When your documents are long, that matters.
The Real Advantage: You Own It #
The benchmark win gets the headline, but the Apache 2.0 license is the argument that closes deals in enterprise. With Voyage, OpenAI, or Cohere, your data goes to their servers — full stop. For legal teams, healthcare systems, and any organization under data residency requirements, that is a dealbreaker regardless of benchmark scores.
With Nemotron 3 Embed, you deploy on your own infrastructure. NVIDIA provides a NIM microservice for production-scale serving, available via build.nvidia.com if you prefer to skip the GPU management overhead. Inference partners — Baseten, FriendliAI, DeepInfra, and OpenRouter — are already running it.
NVIDIA also published the full fine-tuning recipe built on the NeMo framework. Domain adaptation results are concrete: fine-tuning the 1B model on NVIDIA’s documentation corpus improved NDCG@10 from 56.7% to 63.3%. If your retrieval corpus has specialized vocabulary — medical terms, legal citations, internal product names — fine-tuning is now a documented, accessible path.
Who Is Already Running It #
The early signals from agent infrastructure teams are worth attention. Mem0 evaluated the 1B model against their own memory stack and reported 80.38% on LongMemEval Retrieval@10, outperforming larger models. Zep, which provides memory layers for AI agents, ran benchmarks across memory retrieval tasks and had the 1B model finish first against models with significantly more parameters. Zoom and Automation Anywhere are evaluating it for enterprise agentic search.
Agent memory providers moving first makes sense. Embedding quality is the invisible tax on every agentic workflow: if your retriever returns stale or irrelevant context, the agent wastes turns searching again. Better retrieval means fewer iterations, which means lower token costs per task.
Getting Started #
The 1B-BF16 model loads through SentenceTransformers with a two-line install:
pip install --upgrade "transformers>=5.2.0" "sentence-transformers>=5.4.1"
python
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("nvidia/Nemotron-3-Embed-1B-BF16")
queries = ["query: how does RAG work?"]
passages = ["passage: RAG retrieves relevant documents before generation."]
q_emb = model.encode(queries, normalize_embeddings=True)
p_emb = model.encode(passages, normalize_embeddings=True)
scores = q_emb @ p_emb.T # dot product equals cosine similarity (L2-normalized)
The 8B model card on Hugging Face includes examples for Transformers and vLLM. NVIDIA’s official benchmark writeup covers the RTEB methodology and per-task breakdowns for validating fit against your specific retrieval workload.
Open source has been closing the gap on proprietary embeddings for two years. Nemotron 3 Embed is the point where it crossed over.