The collection has three checkpoints for different accuracy and serving tradeoffs across search, RAG, agent memory, and code retrieval.
The headline result is that nvidia/Nemotron-3-Embed-8B-BF16 ranks #1 overall on RTEB with a score of 78.5. RTEB evaluates multilingual retrieval quality across a range of tasks.
The short version:
-
nvidia/Nemotron-3-Embed-8B-BF16 is the accuracy-first option.
-
nvidia/Nemotron-3-Embed-1B-BF16 scores 72.4 on RTEB in a much smaller deployment footprint.
-
nvidia/Nemotron-3-Embed-1B-NVFP4 is optimized for Blackwell, delivering up to 2x BF16 throughput while retaining 99%+ of BF16 retrieval accuracy.
-
All three support multilingual text and code retrieval with a 32K context window.
-
The weights, datasets, and training, fine-tuning, and distillation recipes are open.
Why you should care care: retrieval quality sets the context an agent gets to reason over. A missed document can turn into more searches, more tokens, and a confident answer built on the wrong evidence. Improving retrieval is often a more useful systems lever than asking the reasoning model to recover from bad context.
For correctness’s sake, a leaderboard is one signal, not a substitute for evaluating on your own corpus, query distribution, index, and latency targets. The model suffix matters too, so please include the exact checkpoint when sharing results.
Resources:
If you test them, I’d love to see the setup, comparison, and failure cases. Real workloads > benchmark victory laps.