NVIDIA Nemotron 3.5 Embed 8B

Embedding modelOpen weightsCard summarized

A versatile embedding model developed by NVIDIA for multimodal and multilingual question-answering retrieval

Source: Nemotron-3.5-Embed-8B-BF16 model card (Hugging Face); text scores from the Nemotron-3-Embed-8B-BF16 model card

Use it via API

Self-host only: no published API price.

Code example Compare

Where to call it

Self-host (weights)nvidia/Nemotron-3.5-Embed-8B-BF16
NVIDIA API catalog (build.nvidia.com)nvidia/nemotron-3.5-embed-8b

Supports

Image input

Full feature support

Good for

Multimodal question-and-answer retrieval over large document corpora
Embedding page images with text, tables, charts and infographics
Semantic search and (agentic) RAG pipelines

Overview

NVIDIA · 02Embed

3.5 Embed 8B

A versatile embedding model developed by NVIDIA for multimodal and multilingual question-answering retrieval

Typical price

—

No published price

Dimensions: 4096
Max input: 32,768 tokens
Modes: text embeddings, image embeddings
Released: 20 Oct 2026

Nemotron 3.5 Embed 8B is NVIDIA's open-weights multimodal, multilingual embedding model for question-answering retrieval. It embeds document pages as images, text, or image + text, and retrieves them from text queries; pages can contain text, tables, charts and infographics. It adds a vision encoder to the frozen Nemotron 3 Embed 8B text model, so text embeddings are identical to that model's. Output is one 4096-d L2-normalized vector per input, with a 32,768-token context. model card: Description, Model Architecture

nvidia/Nemotron-3.5-Embed-8B-BF16

Typical price: —
Dimensions: 4096 (can be truncated from the end, e.g. 1024 or 512; re-normalize after truncating)
Max input: 32,768 tokens (text and image tokens combined)
Languages: 34 languages evaluated for text retrieval, 18 for image retrieval
Weights: Open · OpenMDW License Agreement 1.1
Released: 20 Oct 2026

text embeddings
image embeddings
multimodal embeddings (image + text documents)
visual document retrieval