From the source
NVIDIA expanded its Nemotron Retriever collection with open, commercially available models for embedding, reranking and document extraction, aimed at cutting the tokens agents waste on repeated searches.
- The release covers Nemotron 3.5 Embed in 8B BF16, 1B BF16 and 1B NVFP4 variants, Nemotron 3.5 Rerank 8B BF16, and Nemotron 3.5 Parse at 631M parameters.
- NVIDIA says adding Nemotron 3.5 Rerank cut agentic token cost per query by roughly 14% with the 8B embedder and 13% with the 1B, while improving retrieval accuracy in both setups.
- With the 8B retriever, the reranker scored about 0.9 points higher nDCG@10 than Qwen3-VL-Reranker-8B at similar token cost.
- On ViDoRe v3.1, Nemotron 3.5 Embed 8B BF16 reached 61.17 nDCG@10, which NVIDIA says ranked first among dense retrieval models as of 10/20/2026, an 8.94-point gain over Llama Nemotron Embed VL 1B.
- The 1B model, distilled from the 8B teacher, scores 58.38 nDCG@10 and retains 95.4% of the teacher's score, and on MIRACL Vision the 8B model reaches 72.6 nDCG@10, 4.3 points above Qwen3-VL-Embedding-8B.
