Nemotron 3.5 Rerank 8B
NVIDIA's OpenMDW-licensed 8B reranker for text and document-page images, evaluated in 34 languages for text.
In short. Nemotron 3.5 Rerank re-sorts search results for questions about documents. It reads a text question and a candidate, which can be text, a page image or both, and gives a relevance score. It is an 8B model from NVIDIA that is free to use commercially under the OpenMDW license, and it is meant to follow Nemotron 3.5 Embed.
Summary
NVIDIA released Nemotron 3.5 Rerank 8B on Hugging Face on 20 October 2026, with Nemotron 3.5 Embed and Parse, and describes it as a reranker for multilingual and multimodal question answering over documents that may contain text, page images or both, including charts, tables and infographics. It combines a Pixtral-based vision encoder with a bidirectional Ministral 3 language tower. An 8B reranker derived from Ministral-3-8B-Instruct-2512, which is Apache 2.0, was trained first on text-only data and then on image and image-text data. Each input is one text query and one document of up to 32,768 tokens, images have a default longest edge of 1,120 pixels that can be raised to 1,540, and the output is a raw logit for each pair. The card lists 34 languages evaluated for text and 18 for images, and says others may work.
NVIDIA says the model reached the top of the ViDoRe v3.1 reranking leaderboard on 20 October 2026 with a score of 60.49, against 59.17 for Qwen3-VL-Reranker-8B, and that adding it to an agent pipeline cut token use by about 14%. Those figures come from NVIDIA's blog and are the developer's own, and the gap to the Qwen model is small.
The license is OpenMDW-1.1, which allows commercial use, and the BF16 weights are 16.76 GB for 8.38B parameters. The card's examples use Transformers 5.20 with FlashAttention 2, need a CUDA GPU and trust_remote_code, and score documents one at a time. In the 30 days to 11 November 2026, Hugging Face counted about 7,700 downloads.
Good for
- Vision RAG pipelines that search page images with charts and tables, reranking candidates from Nemotron 3.5 Embed or another first stage
- Multilingual question answering over documents, with 34 languages evaluated for text
- Commercial products, thanks to the permissive OpenMDW license
- Reranking text-only candidates too, since the first training stage used text only
Watch out for
- The 16.76 GB BF16 file needs a large GPU
- NVIDIA's lead over Qwen3-VL-Reranker-8B is 1.3 points on one benchmark and its own figure, so test on your documents
- Queries must be text, and every candidate is scored separately, so cost grows with the list
- The examples pin Transformers 5.20.0 and use FlashAttention 2, so plan a dedicated environment
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Nemotron 3.5 Rerank 8B. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Rerank-8B (BF16) BF16 | 8.38B | 17 GB | 19 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.
See also
Frequently asked questions
- What is Nemotron 3.5 Rerank 8B?
- Nemotron 3.5 Rerank re-sorts search results for questions about documents. It reads a text question and a candidate, which can be text, a page image or both, and gives a relevance score. It is an 8B model from NVIDIA that is free to use commercially under the OpenMDW license, and it is meant to follow Nemotron 3.5 Embed.
- Can I use Nemotron 3.5 Rerank 8B commercially?
- Yes. Nemotron 3.5 Rerank 8B is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Nemotron 3.5 Rerank 8B open source?
- Nemotron 3.5 Rerank 8B is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Nemotron 3.5 Rerank 8B on my own computer?
- Yes, on a good graphics card. The smallest variant, Rerank-8B (BF16), needs about 19 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Nemotron 3.5 Rerank 8B released?
- Nemotron 3.5 Rerank 8B was released by NVIDIA on October 20, 2026.
- How long a context window does Nemotron 3.5 Rerank 8B have?
- Nemotron 3.5 Rerank 8B supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Nemotron-3.5-Rerank-8B-BF16 model card (Hugging Face) huggingface.co
- NVIDIA Nemotron 3.5 Embed, Rerank, and Parse (Hugging Face blog, 21 Oct 2026) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Nemotron 3.5 Embed NVIDIA | 1.55B – 8.38B | 32K | OpenMDW-1.1 (Permissive) | 1.46 GB Laptop or phone | |
| Qwen3-VL-Reranker Alibaba Qwen | 2.13B – 8.77B | 32K | Apache 2.0 (Permissive) | 4.9 GB Laptop or phone | |
| Qwen3-Reranker Alibaba Qwen | 596M – 8.19B | 32K | Apache 2.0 (Permissive) | 1.37 GB Laptop or phone | |
| Jina Reranker v3.5 Jina AI | 597M | 128K | CC BY-NC (Non-commercial) | 0.46 GB Laptop or phone |