NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB

Retrieval is the unglamorous plumbing behind every RAG system, and it's usually the thing that quietly breaks when an agent hallucinates or fetches the wrong document. NVIDIA's answer is Nemotron 3 Embed, a family of three open embedding checkpoints built to sit at exactly that layer, handling everything from enterprise search to code lookup to long-term agent memory.

The lineup has a clear hierarchy. Nemotron-3-Embed-8B-BF16 is the accuracy flagship, and as of July 17, 2026 it sits at #1 on RTEB, the Retrieval Embedding Benchmark, scoring 78.46 average NDCG@10 across 16 public tasks. Below it sit two 1B-class models sharing the same architecture: a standard BF16 version and an NVFP4 variant tuned for Blackwell GPUs. All three are transformer encoders built on Mistral's Ministral-3 line — the 8B rides on Ministral-3-8B-Instruct-2512, the two 1B models on Ministral-3-3B-Instruct-2512 — and every checkpoint handles sequences up to 32,768 tokens across 34 languages.

What's notable is how the 1B model actually got made. It isn't a smaller model trained from scratch; it's a compressed descendant of a 3B parent. NVIDIA pruned that 3B down to 2B using ModelOpt's mcore_minitron neural architecture search, picking the best candidate off a Pareto front of ten using a 50,000-sample calibration set, then distilled it against the fine-tuned 8B teacher with a combined cosine-distance and MSE loss. Repeat the process once more and you land at the final 1.14B model. The payoff shows up in the numbers: that 1B checkpoint beats NVIDIA's own prior-generation llama-nemotron-embed-vl-1b-v2 by more than 10 RTEB points.

Then there's the NVFP4 variant, which pushes compression into the serving format itself, quantizing weights and activations in linear layers down to 4 bits and using quantization-aware distillation to claw back accuracy lost on long inputs. The result loses just 0.38 RTEB points versus its BF16 parent — 99.5% retention — while NVIDIA claims up to 2x the throughput on Blackwell hardware. It also supports slicing the 2048-dimension embedding down to 1024 or 512 dimensions if you re-normalize afterward, which is a handy knob for teams trying to shrink index storage without retraining anything.

All three checkpoints ship under the OpenMDW-1.1 license alongside a Rust-based NIM microservice for the 1B model and NeMo AutoModel recipes for fine-tuning. NVIDIA is clearly pitching a two-tier deployment pattern: run 1B-NVFP4 cheaply for bulk recall, then route the harder queries to the 8B model when precision actually matters.