NVIDIA released a new Nemotron 3.5 retrieval suite: Embed (8B and 1B variants, plus a 1B NVFP4 quantised version), a Rerank 8B cross-encoder, and a 631M-parameter Parse model. The 8B embedder is built on Ministral-3-8B-Instruct and trained with contrastive learning for multimodal retrieval, while the 1B model is distilled from it; the reranker reuses the embedder’s backbone with an added MLP scoring head. On the ViDoRe v3.1 benchmark, the 8B embedder scores 61.17 nDCG@10 and the reranker scores 60.49 on combined image-and-text inputs, ahead of Qwen3-VL-Reranker-8B’s 59.17. The Parse model reaches 75.20 on ParseBench while processing 10.5 pages per second on a single H100 GPU in its triple-pass mode.
