Nemotron 3.5 Rerank 8B

Not yet independently verified. Rests on NVIDIA's own model card, which gives 20 October 2026 as the Hugging Face release date; no independent evaluation was found, so no score is stored. We will update this when it can be confirmed, and remove this note.

Nemotron 3.5 Rerank 8B is an open-weights reranking model that NVIDIA released on Hugging Face on 20 October 2026 for multilingual and multimodal question-answering retrieval. Given a text query and a document, which may be text, a page image or both, it returns a score for how relevant the document is.

What it does

A reranker is the second stage of a retrieval system: an embedding model fetches a set of candidate documents, and the reranker reads each one against the query in more detail and reorders them, which improves the precision of what is passed to a language model. NVIDIA says the model handles pages with charts, tables and infographics, and fits classical and agentic vision RAG pipelines. It combines a Pixtral-based vision encoder with a bidirectional language-model tower derived from Ministral 3 8B, trained first on text and then on image and image-text data. Text retrieval was evaluated across 34 languages and image retrieval across 18.

Who it is for

The model suits teams building search or question answering over document collections that include scanned pages and visual material, who want to run the reranker on their own GPUs. NVIDIA states it is ready for commercial use. The default image size is 1,120 pixels on the longest edge, adjustable to 1,540, and each input is one query paired with one document.

Key facts

Maker: NVIDIA Released: 20 October 2026 (Hugging Face) Size: 8.38B parameters Context: 32,768 tokens Inputs: text query with a text, image or mixed document Licence: OpenMDW-1.1 (base model under Apache 2.0)