Qwen3.8-27B NVFP4

Qwen3.8-27B NVFP4

NVIDIA's 4-bit NVFP4 build of Alibaba's Qwen3.8-27B, made with Model Optimizer for Blackwell GPUs, taking text, image and video in and giving text back.

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: Qwen3.8-27B NVFP4 is a compressed copy of Alibaba's Qwen3.8-27B model, made by NVIDIA. It uses a mixed NVFP4 and FP8 format and was produced with the Model Optimizer tool. The card says NVIDIA does not own or develop the underlying model. It only packaged the quantized version.

The model reads text, images and video, and it answers in text. The context length reaches 262K. The record lists 18.2B parameters and about 18.2 GB of files.

Who it suits: It suits developers who want a ready-made quantized model for an agent, a chatbot or a search-and-answer app. It runs on NVIDIA Blackwell GPUs.

Industry | Who uses it | How they use it | Example
Software | App developers | Serve a pre-quantized model behind an app | A chatbot that also reads screenshots
Enterprise IT | Platform teams | Run an internal assistant on Blackwell hardware | A document question tool for staff
AI agents | Agent builders | Use it as the reasoning model in an agent | A task agent that reads images
Research | ML engineers | Test 4-bit serving on a 27B model | Comparing speed with the full model

Skip it if you have no NVIDIA Blackwell GPU, since the card names that as the supported hardware.

Where to use it:

• vLLM, one of the two engines the card names
• SGLang, the other engine the card names
• Linux, which the card gives as the preferred system
• Windows, macOS or Linux, with Python, for the file download alone

Use case | Who | What they get back
Chatbot | Developers | Text answers, with images and video as input
Retrieval apps | Engineers | A reader that handles long context up to 262K
Agent systems | Agent builders | A reasoning model that is already quantized

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

• Python and pip, for the download command
• An NVIDIA Blackwell GPU with enough memory for about 18.2 GB of weights
• Linux, plus vLLM or SGLang as the serving engine

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Start your serving engine and point it at the downloaded folder.

pip install -U huggingface_hub
hf download nvidia/Qwen3.8-27B-NVFP4

How to use it: You serve the model with vLLM or SGLang and send chat requests. The visible part of the card gives no run commands, so follow each engine's own guide for its settings.

Piece | What it is
Input | Text, images and video
Output | Text
Context | Up to 262K
Quantization | Mixed NVFP4 and FP8, via Model Optimizer

Example: You start the model in vLLM on a Blackwell machine and send it a screenshot with a question about the page. The reply comes back as text. You then compare the speed and answers with the full Qwen3.8-27B on the same prompt.

For newcomers and experts: Newcomers should know that quantizing shrinks a model so it fits on less memory and runs faster, with a small quality tradeoff. Begin with a single short prompt and check the answer.

The calibration step used NVIDIA's Nemotron post-training data. The training data of the base model is listed as undisclosed. The card names several evaluation sets, but the visible text gives no scores, so read the original Qwen card for results.

Start from Qwen3.8-27B
Quantize with Model Optimizer
Calibrate on Nemotron data
Serve with vLLM or SGLang

Good to know:

• License: Apache 2.0
• Cost to run: none named, but you need a Blackwell GPU
• Data it reaches: it downloads files from Hugging Face onto your machine
• Not for: GPUs other than Blackwell, since the card lists only that family, and claims of benchmark scores, since none are visible in the card text

On Hugging Face, 13 November 2026

• Task: text generation
• Library: Model Optimizer
• Parameters: 18.2B
• Weights: about 18.2 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 268,673
• Likes: 239
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 5
• Tags: safetensors, qwen3_5, nvidia, modelopt, qwen3.8, quantized, fp4, fp8, text-generation, conversational, 8-bit

Links

• Model: https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4
• Files: https://huggingface.co/nvidia/Qwen3.8-27B-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/Qwen3.8-27B-NVFP4

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