Nvidia ships Nemotron 3.5, open models built for RAG retrieval
Embedding, reranking and document-parsing models from 631M to 8.38 billion parameters, released under Nvidia's OpenMDW licence with benchmark claims to match.
Nvidia released Nemotron 3.5 Retriever on 21 October: three open model families built for retrieval-augmented generation, covering embedding, reranking and document parsing.
Nemotron 3.5 Embed ships in three sizes: an 8.38-billion-parameter BF16 model with a 32,768-token context and 4,096-dimensional embeddings that can be truncated for smaller indexes, plus distilled 1B BF16 and 1B NVFP4 quantised variants for cheaper serving. A matching Rerank-8B cross-encoder scores relevance between a query and its results, and a separate 631-million-parameter Parse model pulls text, tables and charts out of documents. Nvidia’s own benchmarks put Embed-8B at 61.17 nDCG@10 on ViDoRe v3.1, which it says is first among dense retrieval models, and Parse at 75.20 on ParseBench, first among models under 1 billion parameters. Everything ships under Nvidia’s OpenMDW-1.1 licence, with the base model under Apache 2.0, and is available on Hugging Face, through Nvidia NIM, and via hosting partners Baseten, DeepInfra and FriendliAI.
Why it matters: the small variants, a 1B quantised embedder and a 631M parser, are light enough to self-host on a single GPU, while Nvidia’s numbers claim to beat larger retrieval models on the same benchmarks. Anyone building RAG over mixed text-and-image documents, PDFs, slide decks, scanned forms, now has a concrete, permissively licensed model to benchmark against whatever they’re running today.
The caveat: every score above comes from Nvidia’s own benchmark runs; there’s no independent leaderboard entry yet to check them against.