NVIDIA Releases Nemotron 3.5 Embed, Rerank and Parse Retrieval Models
A source-linked abstract prepared once when this article was collected.
NVIDIA announced Nemotron 3.5 Embed, Rerank and Parse, open and commercially available models for embedding, reranking and document extraction. The source reports that Nemotron 3.5 Embed 8B BF16 scores 61.17 nDCG@10 on ViDoRe v3.1, ranking first among dense retrieval models. In agent pipeline experiments, adding Nemotron 3.5 Rerank reduced agentic token cost per query by roughly 13-14% while improving retrieval accuracy. The 1B embedding model is distilled from the 8B teacher and offers NVFP4 quantization for throughput.
Key takeaways
- Nemotron 3.5 Embed 8B BF16 achieves 61.17 nDCG@10 on ViDoRe v3.1, ranking first among dense retrieval models.
- Adding Nemotron 3.5 Rerank reduced agentic token cost per query by about 14% with the 8B embedder and 13% with the 1B embedder.
- Nemotron 3.5 Embed 1B BF16 scores 58.38 nDCG@10, retaining 95.4% of the 8B teacher's score.
- Nemotron 3.5 Embed 1B NVFP4 delivers 27.1% higher throughput than the compared configuration on NVIDIA GB200.
Why it matters
Retrieval quality shapes how many tokens agents spend searching and reasoning. Open, commercially usable multimodal retrievers and rerankers could lower inference costs for enterprise agent workflows that rely on documents containing tables, charts and images.
Read with care
Benchmark results are reported by NVIDIA on its own evaluations and leaderboards; source claims not independently checked.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.