Nemotron 3.5 Embed
NVIDIA's multimodal retrieval models: search document page images and text with 8B or 1B embeddings, plus a matching 8B reranker.
In short. Nemotron 3.5 Embed turns text, and pictures of document pages with their tables and charts, into numbers, so you can search them with an ordinary question. It comes in 8B and 1B sizes, and NVIDIA also released a matching 8B model that re-sorts the best results. The license lets companies use it commercially.
Summary
NVIDIA released Nemotron 3.5 Embed on Hugging Face on 20 October 2026, and announced it on 21 October in a blog post together with Nemotron 3.5 Rerank and Nemotron 3.5 Parse. It is a retrieval model for document pages. Queries are text, while documents can be a page image, text or both, so tables, charts, infographics and schematics can be searched without running OCR first. There are three checkpoints: an 8B model (8.38B parameters, 4,096-dimensional vectors, about 16.8 GB in BF16), a 1B model distilled from it (1.55B parameters, 2,048 dimensions, about 3.1 GB) and an NVFP4 build of the 1B (about 1.3 GB). The vision encoder comes from Ministral 3, which is Apache 2.0, and the language part comes from NVIDIA's text-only Nemotron 3 Embed models and stays frozen. Text embeddings are therefore identical to Nemotron 3 Embed, so existing text indexes can be reused. The context is 32,768 tokens, and vectors can be cut to 1,024 or 512 dimensions and renormalized.
NVIDIA reports 61.17 nDCG@10 for the 8B model on ViDoRe v3.1, first among dense retrieval models on that leaderboard on 20 October 2026, against 52.23 for its earlier Llama Nemotron Embed VL 1B. The 1B model scores 58.38, which is 95.4% of the 8B result. On MIRACL Vision the 8B model scores 72.6, 4.3 points above Qwen3-VL-Embedding-8B. NVIDIA says the NVFP4 build keeps 99.85% of the 1B model's ViDoRe v3 accuracy while raising throughput by 27.1% on a GB200 GPU. The companion Nemotron 3.5 Rerank 8B scores 60.49 on the ViDoRe v3.1 reranking leaderboard against 59.17 for Qwen3-VL-Reranker-8B, and NVIDIA says adding it cuts agent token use by about 14%. These are all NVIDIA's figures.
All of the models use the OpenMDW-1.1 license, and the cards say they are ready for commercial or non-commercial use. The cards give usage examples for Transformers, Sentence Transformers and vLLM, and NVIDIA also offers NIM containers for serving. Text retrieval was evaluated in 34 languages and image retrieval in 18, and the cards say other languages may work but do not claim them.
Good for
- Searching scans, slides and reports by their page images, including tables, charts and infographics, without an OCR step
- Search systems that already use Nemotron 3 Embed, since text embeddings stay identical and indexes can be reused
- Smaller deployments with the 1B model (about 3.1 GB) or its NVFP4 build (about 1.3 GB)
- Two-stage search with the matching Nemotron 3.5 Rerank 8B
Watch out for
- Queries must be text, so you cannot search with an image
- The cards only claim evaluation in 34 languages for text and 18 for images
- The NVFP4 speed gain was measured on NVIDIA GB200 hardware
- Every score is NVIDIA's own, and the leaderboard positions are dated 20 October 2026
- The 8B vectors need 4,096 dimensions of storage each unless you shorten them
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 Embed. 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 |
|---|---|---|---|---|
| Embed-1B (BF16) BF16 | 1.55B | 3.11 GB | 3.58 GB | Laptop or phone |
| Embed-1B (NVFP4) NVFP4 | 1.55B | 1.27 GB | 1.46 GB | Laptop or phone |
| Embed-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
- Best open-source embedding models in 2026
- EmbeddingGemma 2 embeds text, images, video and audio under Apache 2.0
- NVIDIA's Nemotron 3.5 Embed and Parse: document pages as images
Frequently asked questions
- What is Nemotron 3.5 Embed?
- Nemotron 3.5 Embed turns text, and pictures of document pages with their tables and charts, into numbers, so you can search them with an ordinary question. It comes in 8B and 1B sizes, and NVIDIA also released a matching 8B model that re-sorts the best results. The license lets companies use it commercially.
- Can I use Nemotron 3.5 Embed commercially?
- Yes. Nemotron 3.5 Embed 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 Embed open source?
- Nemotron 3.5 Embed 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 Embed on my own computer?
- Yes. The smallest variant, Embed-1B (NVFP4), needs about 1.46 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Nemotron 3.5 Embed released?
- Nemotron 3.5 Embed was released by NVIDIA on October 20, 2026.
- How long a context window does Nemotron 3.5 Embed have?
- Nemotron 3.5 Embed 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.
- NVIDIA Nemotron 3.5 Embed, Rerank, and Parse (Hugging Face blog, 21 Oct 2026) huggingface.co
- Nemotron-3.5-Embed-8B-BF16 model card (Hugging Face) huggingface.co
- Nemotron-3.5-Embed-1B-BF16 model card (Hugging Face) huggingface.co
- Nemotron-3.5-Rerank-8B-BF16 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-VL-Embedding Alibaba Qwen | 2.13B – 8.14B | 32K | Apache 2.0 (Permissive) | 4.9 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone | |
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone |