Reflection Beam 1.0
An open-weight text model for coding, reasoning, and tool-using applications, available through hosted and self-hosted deployment.
NVIDIA Nemotron 3.5 Rerank 8B
huggingface.co
A multilingual reranking model that scores text and document images for relevance to a text query.
Visit website
Pricing
Free
Available on
Self-hosted, NVIDIA CUDA, Hosted API
Category
Developer Tools
What is NVIDIA Nemotron 3.5 Rerank 8B?
NVIDIA Nemotron 3.5 Rerank 8B evaluates candidate documents against a text query. Its BF16 checkpoint handles text, document images, and combined inputs, including charts, tables, and infographics.
It serves search and retrieval-augmented generation developers who need to reorder results after an initial retrieval step. NVIDIA also documents fine-tuning for collections where relevant candidates are retrieved but placed below less relevant material.
Key features
Document relevance scores
Returns raw logits for query-document pairs; higher values indicate greater predicted relevance.
Multilingual evaluation coverage
NVIDIA evaluated text retrieval across 34 languages and image retrieval across 18 languages.
Hosted passage ranking
The API accepts 1–512 passages per request, with text, an image, or both in each passage.
Domain adaptation workflow
Fine-tuning guidance supports text-only, image-only, and paired documents for adapting candidate ordering.
Who is it for?
RAG evidence selection
Reorder candidate sources before passing selected evidence to an answer generator.
Visual document search
Prioritize retrieved report pages containing relevant charts or tables.
Specialist collection adaptation
Explore fine-tuning for a document collection with domain-specific relevance judgments.
How to use NVIDIA Nemotron 3.5 Rerank 8B
- Choose the downloadable BF16 checkpoint or the hosted reranking endpoint.
- Retrieve an initial candidate set from your search system.
- Prepare a text query and candidate documents using the selected interface's input format.
- Score the candidates and order them by relevance.
- Evaluate the resulting ranking on representative queries before integration.
NVIDIA Nemotron 3.5 Rerank 8B pricing
Free
Downloadable model materials are available without a license fee under OpenMDW-1.1. This covers the model license; compute and hosted service costs are separate.
Check current pricing ↗What to know before you use it
- Reranking operates on retrieved candidates. When relevant documents are missing from that set, improve first-stage retrieval before adapting the reranker.
- The downloadable model supports 32,768 combined text and image tokens; the hosted endpoint instead limits each query-passage pair to 10,240 tokens.
- Oversized hosted pairs are rejected by default. END truncation removes passage text while preserving the query and image.
Frequently asked questions
Does it generate answers?
Its documented output is relevance scores for documents. An answer generator is a separate pipeline component.
Can it rank image-only documents?
Yes. Documents may contain an image alone, text alone, or both; the query remains text.
When should I consider fine-tuning?
NVIDIA recommends the workflow when relevant candidates reach the retrieved set but rank below less relevant results.