NVIDIA · 2026-07-16 · major
Nemotron 3 Embed — NVIDIA's open 8B embedder takes #1 on RTEB
NVIDIA released the Nemotron 3 Embed family — 8B and two 1B open embedders with 32K context. The 8B model hits 78.5% on RTEB and 75.5% on MMTEB Retrieval, ranking first overall on the RTEB leaderboard.
NVIDIA ships an open 8B embedder that takes #1 on RTEB, plus two efficient 1B variants for production RAG.
Key specs
| Context window | 32K tokens |
|---|---|
| Rteb (8 b) | 78.5% |
| Mmteb retrieval (8 b) | 75.5% |
| Params (flagship) | 8B |
Quick facts
| Maker | NVIDIA |
|---|---|
| Family | Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, Nemotron-3-Embed-1B-NVFP4 |
| Base model | Adapted from Ministral-3-8B-Instruct-2512 (Apache-2.0) |
| License | OpenMDW-1.1 |
| Context window | 32,000 tokens |
| RTEB (8B) | 78.5% — rank #1 overall |
| Availability | Hugging Face, NVIDIA NIM, Baseten, DeepInfra, Friendli AI, OpenRouter |
Benchmarks
What is it?
Nemotron 3 Embed is a family of three open embedding models from NVIDIA aimed at retrieval-augmented generation, agentic retrieval, and agent memory. The flagship Nemotron 3 Embed 8B ranks first overall on RTEB with 78.5%, and NVIDIA also ships a 1B BF16 model for production and a 1B NVFP4 variant tuned for Blackwell GPUs. All three take a 32K-token input window.
How does it work?
The 8B model adapts Mistral's Ministral-3-8B-Instruct backbone by turning the causal decoder into a bidirectional encoder and fine-tuning on contrastive pre-training over web and synthetic text pairs, followed by curated multilingual retrieval data across legal, finance, medical, business, and education domains. The two 1B variants are distilled and pruned from a 3B retriever built on Ministral-3-3B using NVIDIA's ModelOpt NAS engine, with two-stage progressive context scaling from 1024 to 4096 tokens before the full 32K extension.
Why does it matter?
Retrieval quality is the ceiling on how good any RAG or agent pipeline gets, and Nemotron 3 Embed 8B setting a new #1 on RTEB with open weights, published training data, and released fine-tuning recipes gives every team a credible open alternative to closed embedders. The 1B NVFP4 variant plus vLLM support means production inference costs are within reach on a single Blackwell GPU.
Who is it for?
Teams building enterprise RAG, agent memory, and code retrieval; researchers who need reproducible SOTA retrieval baselines.
Frequently asked questions
- How does Nemotron 3 Embed 8B compare on retrieval benchmarks?
- Nemotron 3 Embed 8B posts 78.5% on RTEB, taking the #1 slot on the RTEB leaderboard for retrieval accuracy across real-world tasks. It also scores 75.5% on MMTEB Retrieval and 64.4 on LMEB. The 1B BF16 variant reaches 72.4% on RTEB and 71.0% on MMTEB Retrieval, with NVIDIA claiming ~27% error reduction over the previous generation.
- What is the license on Nemotron 3 Embed?
- Nemotron 3 Embed ships under the OpenMDW-1.1 (Open Model, Data, and Weights) license from NVIDIA. The underlying Ministral-3-8B-Instruct-2512 backbone that Nemotron 3 Embed adapts is Apache-2.0. NVIDIA also publishes the training data, recipes, and NeMo AutoModel fine-tuning and distillation scripts alongside the weights.
- Where can developers deploy Nemotron 3 Embed?
- Nemotron 3 Embed is available day 0 on Hugging Face for direct download, as an NVIDIA NIM microservice, and through cloud partners including Baseten, DeepInfra, Friendli AI, and OpenRouter. The 1B-NVFP4 variant is optimized for Blackwell GPUs, and vLLM support ships from launch.
- How was the 1B Nemotron 3 Embed trained?
- The 1.14B Nemotron 3 Embed was not trained from scratch. NVIDIA started from a 3B retriever base built on Ministral-3-3B, compressed it to 2B using its ModelOpt NAS engine, then applied structured pruning and distillation from the 8B teacher across two compression cycles. Context was scaled progressively from 1024 to 4096 tokens before extending to the full 32K window.
Try it
https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16