Nemotron 3.5 Embed 8B

Nemotron 3.5 Embed 8B is an open-weights embedding model that NVIDIA released on Hugging Face on 20 October 2026 for multimodal and multilingual question-answering retrieval. It turns a text query, or a document page given as text, an image or both, into a vector for semantic search.

## What it does

The model is a bi-encoder: a whole corpus of page images or text chunks is embedded once and stored in a vector database, and each query is embedded at search time and matched against it. Token-level states are mean-pooled and normalised into one 4,096-dimension vector per input, and the vector can be truncated from the end, for example to 1,024 or 512 dimensions, to save storage, provided it is normalised again. Queries are text only; documents may be text, images or both, including pages with tables, charts and infographics. Text retrieval was evaluated across 34 languages and image retrieval across 18.

## Who it is for

It is the first stage of a retrieval pipeline and pairs with NVIDIA's Nemotron 3.5 Rerank 8B, which reorders the candidates it returns. Teams building search or RAG over large document collections with scanned or visual pages, who want to host the model themselves, are the intended users, and NVIDIA states it is ready for commercial use.

## Key facts

**Maker:** NVIDIA **Released:** 20 October 2026 (Hugging Face) **Size:** 8.38B parameters **Context:** 32,768 tokens **Output:** 4,096-dimension embedding, truncatable **Licence:** OpenMDW-1.1 (base model under Apache 2.0)

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.