NVIDIANemotron 3.5
Nemotron 3.5 Rerank 8B
This 8.38B model scores a text query against a candidate document. The document can be text, an image or a combination.
What the release provides
It uses a Pixtral-based vision encoder and a bidirectional Ministral 3 language tower derived from Ministral 3 8B Instruct. Its maximum sequence length is 32,768 tokens.Nemotron 3.5 Rerank 8B official model card
How to use it well
Practical guidance. Run it after a first retrieval step has produced a manageable candidate set. Keep the retrieved evidence even when the reranker lowers its position. Tune the candidate count against answer coverage and latency on real queries.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | nvidia/Nemotron-3.5-Rerank-8B-BF16 |
| Release license | OpenMDW 1.1; Ministral base under Apache 2.0Nemotron 3.5 Rerank 8B official model card |