Skip to main content
Watch: I Built an AI App Without Leaving Databricks

NVIDIA - Open weights - verified

Nemotron 3.5 Embed 8B embeddings

Nemotron 3.5 Embed 8B is a current embedding model from NVIDIA, with open weights on Hugging Face at nvidia/Nemotron-3.5-Embed-8B-BF16. It returns 4,096-dimension vectors by default and accepts up to 32,768 input tokens, for text and image input. There is no API price: the weights download from Hugging Face under OpenMDW 1.1, and you pay for the compute you run them on.

CurrentModel id: nvidia/Nemotron-3.5-Embed-8B-BF16- weights: nvidia/Nemotron-3.5-Embed-8B-BF16

Default dims
4,096
Max input
32,768 tok
Price / 1M
Self-hosted
Inputs
Text, Image

Dimensions

Flexible output size
  • 4,096default
  • 1,024
  • 512

Default output: 4,096 dimensions.

Listed sizes: 4,096, 1,024 and 512.

NVIDIA documents shorter output vectors for this model (Matryoshka-style), so you can store smaller embeddings by asking for fewer dimensions instead of re-embedding with another model.

Pricing

No API price: the weights are on Hugging Face and you run them on your own hardware or host.

Open weights

License
OpenMDW 1.1openmdw-1.1
Commercial use
Commercial use allowedCommercial use is allowed, per the license.
Parameters
8.38B (8,381,155,328)

Worth knowing

Queries are text only with the `query:` prefix while documents can be text, page images or both with `passage:`, and the card lists a Hugging Face release date of 10/20/2026 although the repo was created 2026-10-07.

Released . Multilingual: yes.

Usage

Copied verbatim from the page linked below. Placeholders such as project ids and API keys are the source's own.

Terminal
vllm serve "$MODEL_PATH" \
    --served-model-name "nvidia/Nemotron-3.5-Embed-8B-BF16" \
    --dtype bfloat16 \
    --max-model-len 32768 \
    --chat-template-content-format openai \
    --host 127.0.0.1 \
    --port 8000

Source: huggingface.co/nvidia/Nemotron-3.5-Embed-8B-BF16

FAQ

How many dimensions does Nemotron 3.5 Embed 8B output?

Default output: 4,096 dimensions. Listed sizes: 4,096, 1,024 and 512. NVIDIA documents shorter output vectors for this model (Matryoshka-style), so you can store smaller embeddings by asking for fewer dimensions instead of re-embedding with another model.

What is the max input length of Nemotron 3.5 Embed 8B?

32,768 tokens per input, per the model card. Longer text has to be chunked before you embed it.

Can I use Nemotron 3.5 Embed 8B commercially?

Nemotron 3.5 Embed 8B is released under OpenMDW 1.1, and commercial use is allowed. The license text is at https://openmdw.ai/license/1-1/.

What can Nemotron 3.5 Embed 8B embed?

Text and image input, per the provider. It is documented as multilingual.

Sources

Last verified from the pages below.

Other open-weight embedding models

Default dimensions and max input tokens for the other models you can run yourself.

All embedding models - Open-weight embedding models

Newsletter

New models, every week

One email per week. New model releases, embedding and retrieval updates and what they mean for what you build.

Keep exploring

Model API pricing and specs

First-party prices per 1M tokens for chat and reasoning models.

Local models

Open-weight LLMs you can download and run, with sizes and licenses.

What is RAG?

Where embeddings fit in a retrieval pipeline, end to end.

Vector databases compared

Where the vectors go once you have picked a model.