DeepSeek-V4.1-Flash-NVFP4

DeepSeek-V4.1-Flash-NVFP4

Reads text and images and writes text, as a smaller NVFP4 copy of DeepSeek-V4.1-Flash made by NVIDIA. For developers deploying agents, chatbots and retrieval systems on NVIDIA Blackwell GPUs.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is NVIDIA's quantized copy of DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts language model that takes text and images in and writes text out. NVIDIA made it with Model Optimizer. Context can reach up to one million tokens.

Who it suits: Developers who want a pre-quantized model they can deploy as is, instead of quantizing one themselves.

Industry | Who uses it | How they use it | Example
Software | AI engineers | Serve the model behind an agent system | Run an agent that reads screenshots and text
Support | Chatbot builders | Use it as the model inside a chatbot | Answer questions about a pasted image
Search | Retrieval builders | Use it to write answers from retrieved documents | Answer from long documents in one prompt

Small teams without Blackwell GPUs and a lot of disk space can skip it.

Where to use it:

• Linux machines with NVIDIA Blackwell GPUs
• vLLM, but only a build that can load the MXFP4 Engram tables
• SGLang, but only the previous revision of the files

Use case | Who | What they get back
Agent systems | Developers | A model that reads text and images and writes text
Chatbots | Developers | Replies from a large model in a smaller quantized form
Retrieval systems | Developers | Answers written over long context

How to set it up:

• Python with pip, to fetch the files
• A Linux machine with NVIDIA Blackwell hardware
• About 763 GB of disk space for the weights
• A vLLM build that supports the MXFP4 Engram tables, or the previous revision of the files

1. Install the Hugging Face hub tool.
2. Download the model files.
3. Serve them with a supported vLLM build, or pin the previous revision if your build cannot load the current one.

pip install -U huggingface_hub
hf download nvidia/DeepSeek-V4.1-Flash-NVFP4

To pin the previous revision, the card gives this flag.

--revision 3431dde3247c13b5957f682b1e3c6fcae2566079

How to use it: Once the model is served, you send it text, or text with an image, and read back text. The card does not list a serve command, so follow the runtime's own instructions for loading it.

Example: You send a screenshot of a chart with the question of what trend it shows. The model reads the image and the question together and writes a text answer. This is an illustration of the input and output types the card lists, not a tested run.

For newcomers and experts: If you are new, know that this is an enormous model. Downloading it needs hundreds of gigabytes, and the card names only Blackwell GPUs on Linux. A first small step is to read DeepSeek's own model card for the original.

For experts, the card counts 552B parameters in the backbone. The model activates 8B while reading the prompt and 16B while writing, and adds a 196B Engram memory. Weights are stored in NVFP4, and the Engram lookup tables in MXFP4 since the October update.

DeepSeek-V4.1-Flash is the base model
Model Optimizer quantizes it to NVFP4
You download the files from Hugging Face
vLLM serves it on Blackwell GPUs

Good to know: NVIDIA did not build the original model. The card says it was made to a third party's requirements and NVIDIA only quantized it. It also says systems built on it need extra testing with your own data.

• License: MIT, as the Hugging Face record lists it.
• Cost to run: no API key or account is named. You need Blackwell hardware and about 763 GB of weights.
• Data it reaches: downloading reaches Hugging Face. The card does not describe the model reaching files, the web or an account.
• Not for: stock vLLM images cannot load the current revision, and SGLang was not validated with the MXFP4 tables.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: Model Optimizer
• Parameters: 763B
• Weights: about 763 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 33,272
• Likes: 128
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash
• Quantized versions: 1
• Tags: safetensors, deepseek_v41, nvidia, modelopt, deepseekv4, quantized, fp4, image-text-to-text, 8-bit, fp8

Links

• Model: https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-NVFP4
• Files: https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/DeepSeek-V4.1-Flash-NVFP4

Then follow the usage in its model card. Nothing in this kit runs by itself.