Nemotron-3.5-Rerank-8B-BF16
by nvidia
NVIDIA's 8.38B multimodal reranker: scores how relevant a page image, a text passage or both are to a text query, across the languages Nemotron-3.5 covers
nvidia/Nemotron-3.5-Rerank-8B-BF16Deploy Nemotron-3.5-Rerank-8B-BF16
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Nemotron-3.5-Rerank-8B-BF16 is a cross-encoder reranker from NVIDIA, published on Hugging Face in October 2026 under the OpenMDW-1.1 license, which allows commercial use. It reads one text query and one document together and returns a raw relevance logit, where higher means more relevant. A document can be text, a page image (charts and tables included) or both.
It is the second stage of a retrieval pipeline. A first-stage search, such as Nemotron-3.5-Embed-8B over page images, returns a few dozen candidates, and this model rescores each pair so the best pages reach the top before they go to a language model. NVIDIA evaluated it with NDCG@10 on ViDoRe V3 and MIRACL-VISION. The card publishes no scores, and we have not measured it.
Architecture
A Pixtral-based vision encoder and a bidirectional Ministral3 language-model tower (Mistral3ForSequenceClassification in config.json), 8.38B parameters, BF16 weights, derived from Mistral's Ministral-3-8B-Instruct-2512. Training ran in two stages: first on text-only data (21M samples, per the card), then on image and image-text data (2.3M samples) from the text checkpoint. Each input is one query and one document rendered through the model's chat template with query and document roles; the output is one raw logit per pair.
Mixpeek SDK Integration
# Rerank first-stage results with Nemotron-3.5-Rerank-8B-BF16 on your own GPU
# (A100 80GB, H100 80GB or GB200 per the card). A reranker scores at query time,
# so there is nothing to index for it; index pages with your embedder as usual,
# for example Nemotron-3.5-Embed-8B-BF16 into a Mixpeek Vector Store namespace.
import torch
from PIL import Image
from sentence_transformers import CrossEncoder
reranker = CrossEncoder(
"nvidia/Nemotron-3.5-Rerank-8B-BF16",
device="cuda:0",
max_length=32768,
model_kwargs={"attn_implementation": "flash_attention_2", "dtype": torch.bfloat16},
)Capabilities
- Reranks text-only, image-only and image plus text documents against a text query
- Scores pages with charts, tables and infographics from the page image
- Up to 32,768 tokens per query-document pair; images at a 1,120-pixel longest edge by default, up to 1,540
- Runs with Transformers, Sentence Transformers CrossEncoder and vLLM
- OpenMDW-1.1 license; the base model is Apache-2.0
Use Cases on Mixpeek
Performance
The card lists A100 80GB, H100 80GB and GB200 186GB as supported GPUs, Linux only, and its examples use Flash Attention 2 with transformers 5.20.0. A cross-encoder scores every candidate separately, so cost grows with the number of candidates: rerank the top few dozen from a first-stage search. Scores are raw logits (the card's examples are negative numbers), so compare them within one query rather than against a fixed threshold. No benchmark scores are published on the card; we have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is Nemotron-3.5-Rerank-8B-BF16?
An NVIDIA cross-encoder reranker with 8.38B parameters. It takes a text query and one document (text, a page image, or both) and returns a relevance logit, so you can reorder search results before they reach a language model. It is released under the OpenMDW-1.1 license.
What is the difference between Nemotron-3.5-Embed and Nemotron-3.5-Rerank?
The embedder turns each page and each query into a vector once, so it can search millions of pages fast. The reranker reads the query and one page together and scores the pair, which is slower and more precise. Use the embedder to find the top 50 and the reranker to order them.
Can it rerank PDF pages with charts?
Yes. A document can be the page image, so evidence that only appears in a chart or table can still score. The card's examples rank an image plus text document above the same text alone.
What does a Nemotron-3.5-Rerank score mean?
It is a raw logit, and higher is more relevant. The card's example scores are all negative (about -1.8 for the best match, -2.96 for an unrelated one), so use the scores to order candidates for one query. To filter with a threshold, calibrate it on queries you have labelled.
How do I use Nemotron-3.5-Rerank with Mixpeek?
Search a Mixpeek Vector Store namespace for candidates, then rerank them with this model on your own GPU, as in the example on this page. Mixpeek's managed rerank stage ships BAAI/bge-reranker-v2-m3 by default; running this model inside a retriever means registering it as a custom reranker, which is an Enterprise path.
Specification
Build a pipeline with Nemotron-3.5-Rerank-8B-BF16
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free