DeepSeek-V4.1-Flash-1017-NVFP4

DeepSeek-V4.1-Flash-1017-NVFP4

NVIDIA's NVFP4 quantized copy of DeepSeek AI's DeepSeek-V4.1-Flash-1017, a huge multimodal mixture-of-experts model for 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: NVIDIA took DeepSeek AI's DeepSeek-V4.1-Flash-1017 and shrank its weights to NVFP4 with the Model Optimizer tool. This repository holds the result. The base model is a mixture-of-experts model that takes text and images and writes text. It supports a context of up to 1M tokens.

NVIDIA says it did not develop the original model. DeepSeek AI did. The file set is large, about 665 GB.

Who it suits: It suits teams with serious GPU hardware who want a pre-quantized model for agent systems, chatbots or retrieval-augmented generation. The card names NVIDIA Blackwell GPUs and Linux.

Industry | Who uses it | How they use it | Example
AI platforms | Infrastructure teams | Serve a huge model on Blackwell GPUs | Hosting an internal assistant
Software | Application developers | Build agent systems on a pre-quantized model | A tool-calling agent for support tickets
Search and knowledge | RAG teams | Answer questions over long documents with a 1M context | Questions over a large report set
Research | ML engineers | Study how NVFP4 quantization behaves on a large model | Comparing against the original weights

Skip it if you do not have Blackwell class GPUs and several hundred gigabytes of storage.

Where to use it:

• vLLM, in a build that includes the Engram changes the card describes
• SGLang
• NVIDIA Blackwell GPUs on Linux
• Windows, macOS or Linux, with Python, for the file download only

Use case | Who | What they get back
Agent systems | Developers | A model that reads text and images and writes text
Chatbots | Product teams | Replies backed by a long context window
RAG systems | Knowledge teams | Answers grounded in retrieved documents

How to set it up: You need a few things in place first.

• Python and pip, for the download command
• A vLLM build with the MXFP4 Engram support, or SGLang
• NVIDIA Blackwell GPUs on Linux
• About 665 GB of disk space, from the record

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Serve it with a runtime the card lists. Stock vLLM images cannot load the Engram tables, so use a build with the needed changes.

pip install -U huggingface_hub
hf download nvidia/DeepSeek-V4.1-Flash-1017-NVFP4

How to use it: Once the server is running, you send it text or images through its endpoint and read the text it writes back. The model card has no run commands in the part available here. NVIDIA also says a deployment needs testing with your own data before it goes live.

Example: You serve the model on a Blackwell machine and send it an image of a chart with a question about it. The model replies in text. Output is always text, even when the input is an image.

For newcomers and experts: This is a huge model, so most people should try the original through a hosted service first. Download this one only if you have the hardware.

The model has 552B backbone parameters in total, with 8B active during prefill and 16B during decode, plus a 196B Engram conditional memory. NVIDIA quantized it to NVFP4 with Model Optimizer version 0.48.0rc0. It used cnn_dailymail and a Nemotron post-training dataset for calibration. The Engram tables are stored as MXFP4.

Download the files with hf
Serve with a vLLM build or SGLang
Send text or an image
The model writes text back

Good to know:

• License: MIT
• Cost to run: none named, but it needs Blackwell GPUs and about 665 GB of storage
• Data it reaches: it downloads files from Hugging Face onto your machine
• Not for: stock vLLM images, which cannot load this checkpoint

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: Model Optimizer
• Parameters: 665B
• Weights: about 665 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 13,127
• Likes: 188
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• 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-1017-NVFP4
• Files: https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-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-1017-NVFP4

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