Introduction
On July 16, 2026, NVIDIA released Nemotron 3 Embed, a family of open-weight text embedding models aimed at retrieval and retrieval-augmented generation (RAG). NVIDIA reports that the largest model, Nemotron-3-Embed-8B, ranked #1 overall on RTEB — the multilingual Retrieval Text Embedding Benchmark — with a 78.5% score. The release ships with open weights and open training recipes, and a 4-bit variant tuned for NVIDIA's Blackwell GPUs.
Embedding models turn text into vectors so that semantically related passages land close together, which is what makes semantic search and RAG pipelines work. This launch follows NVIDIA's October 2025 Omni-Embed-Nemotron-3B, but where that model targeted multimodal content, Nemotron 3 Embed is focused on text and code retrieval.
The three models
The collection has three variants, all with a 32k-token context window and mean pooling, and all using query: / document: input prefixes:
- Nemotron-3-Embed-8B-BF16 — 8.0B parameters, 4096-dimensional embeddings. The top-performing model.
- Nemotron-3-Embed-1B-BF16 — 1.14B parameters, 2048-dimensional embeddings. The efficiency option.
- Nemotron-3-Embed-1B-NVFP4 — the 1B model in NVIDIA's 4-bit NVFP4 format, quantized for Blackwell.
The NVFP4 build is the headline efficiency claim: NVIDIA reports up to 2x higher throughput than the BF16 model on Blackwell while retaining more than 99% of its retrieval accuracy — the kind of trade-off that matters when you are re-embedding a large corpus or serving retrieval at scale.
Benchmark results
All of the numbers below are NVIDIA-reported and should be read as vendor results until third-party leaderboards confirm them:
- RTEB (multilingual retrieval): 8B scored 78.5% for #1 overall; the 1B model scored 72.4%, which NVIDIA frames as a 27% error reduction over its predecessor.
- MMTEB Retrieval (average NDCG@10): 8B scored 75.5%; 1B scored 71.0% (a reported 28% error reduction).
- LMEB: 8B reached 64.4%, which NVIDIA calls a new state of the art; 1B reached 61.5%.
NVIDIA also cites results on ViDoRe V3 and LongEmbed, and a fine-tuning example on its own documentation where the 1B model's NDCG@10 rose from 56.7% to 63.3% — a reminder that domain fine-tuning still beats an off-the-shelf model on in-domain retrieval.
How NVIDIA built them
The two model sizes were produced differently. The 8B model adapts Ministral-3-8B-Instruct-2512, converting a causal decoder — which only attends to earlier tokens — into a bidirectional encoder better suited to embedding whole passages. The 1B model was compressed from a 3B base through two rounds of structured pruning (via NVIDIA's ModelOpt neural architecture search) and knowledge distillation from the 8B model as teacher. Training was progressive: a first stage at 1024-token context, then a second stage at 4096 tokens with long-context data.
Availability and licensing
NVIDIA released the models with open weights and open-source training recipes. They are available:
- On Hugging Face, with model cards and code examples;
- As an NVIDIA NIM microservice on build.nvidia.com;
- Through cloud partners including Baseten, Bitdeer AI, DeepInfra, Friendli AI, and OpenRouter;
- With vLLM support for self-hosted serving.
Early enterprise users quoted in the announcement — Automation Anywhere, Mem0, and You.com — point to the open weights and fine-tuning recipes, and to gains from swapping the model into existing retrieval and re-ranking stacks.
Why it matters
Retrieval quality is the quiet bottleneck in most RAG and agentic systems: if the embedding model surfaces the right evidence early, the downstream model reads fewer documents and runs fewer follow-up searches, which cuts token cost and latency. An open model that tops a multilingual retrieval leaderboard — and ships a Blackwell-optimized build that doubles throughput at near-parity accuracy — gives teams a credible, self-hostable alternative to proprietary embedding APIs. The caveat is the usual one: the standout numbers are the vendor's own, so treat the leaderboard claim as a strong starting hypothesis and benchmark against your own corpus before committing.
Conclusion
Nemotron 3 Embed is a focused, practical release: three sizes, a genuinely long 32k context, open weights and recipes, and a quantized variant built for the hardware NVIDIA sells. If the RTEB result holds up outside NVIDIA's own evaluation, the 8B model is a strong default for multilingual retrieval, while the 1B NVFP4 build targets teams that care most about throughput. As always with embedding models, the real test is retrieval accuracy on your own documents — start there.
To learn more, explore embedding, vector search, and information retrieval in our glossary, or browse the latest AI models.
Sources
- NVIDIA: Nemotron 3 Embed wins RTEB (Hugging Face blog)
- Nemotron 3 Embed collection on Hugging Face
- Nemotron-3-Embed-8B-BF16 model card
- Nemotron-3-Embed-1B-BF16 model card
- Nemotron-3-Embed-1B-NVFP4 model card
- Nemotron 3 Embed NIM microservice (build.nvidia.com)
- Ministral-3-8B-Instruct-2512 base model