Qwen3.8-27B-NVFP4

Qwen3.8-27B-NVFP4

A 4-bit NVFP4 copy of the Qwen3.8 27B vision-language model made by Unsloth, for serving with vLLM. For developers who want a dense model that reads images and video and works in agent tools.

Made by unsloth 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 Unsloth's NVFP4 quantized version of Qwen/Qwen3.8-27B. The original is a dense 27B model that understands text, images and video, with a switchable thinking mode. The NVFP4 files are Safetensors, and the record lists about 19.9 GB of weights.

Who it suits: Developers who serve large models on their own hardware and want a smaller 4-bit copy of Qwen3.8 for coding and agent work.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a local model behind coding agents such as Codex | A local endpoint with developer role support
Research | AI researchers | Test a quantized vision-language model | Ask questions about a diagram
Enterprise | Platform teams | Serve a private model with vLLM | An internal assistant that reads documents
Media | Video teams | Ask questions about long videos | A summary of an hour of footage

You can skip it if you use SGLang with a version other than 0.5.19, since the card says the FP8 output layer limits it to that one.

Where to use it:

• vLLM, which the card says works with this model
• SGLang, but only version 0.5.19
• Unsloth Desktop, which the card says can run and fine-tune Qwen3.8
• Agent tools such as Codex, through developer role support

Use case | Who | What they get back
Coding and agent tasks | Developers | A model tuned for long, multi-step work
Image and video questions | Analysts | Text answers about pictures and video
Tool calling | Developers | Better parsing of nested objects in tool calls

How to set it up:

• Python and pip on your computer
• A GPU runtime such as vLLM that supports NVFP4
• About 19.9 GB of disk for the weights

1. Install the Hugging Face hub package.
2. Download the model files with the hf command.
3. Serve them with vLLM, following the Unsloth guide named on the card.

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

How to use it: Serve the model with vLLM and send it chat requests. Thinking mode is on by default and you can turn it off per request. Reasoning depth uses the reasoning_effort setting, and preserve_thinking keeps earlier reasoning in the history.

Setting | What it does
reasoning_effort | Tunes how deeply the model reasons
preserve_thinking | Keeps reasoning from earlier messages in the context
thinking mode | Is on by default and can be turned off for a single request
presence_penalty | Can be raised to reduce endless repetition, with a risk of mixed languages

Example: You ask it to plan a refactor of a codebase in thinking mode. It reasons through the steps first and then returns a plan, which an agent tool can carry out.

For newcomers and experts: If you are new, use vLLM and the sampling values the card suggests, and test with a short question first. This is a large model, so check your GPU memory.

If you know the field, the 27B base has 64 layers that alternate Gated DeltaNet and Gated Attention blocks. Context is 262,144 tokens natively. The card says the output layer stays FP8 rather than 4-bit, and that is why SGLang needs version 0.5.19. It also mentions MTP for faster inference.

Qwen3.8-27B base model
Unsloth NVFP4 quantization
Serve with vLLM
Text, image and video answers

Good to know:

• License: Apache 2.0
• Cost to run: your own GPU hardware, with no paid service named
• Data it reaches: the files you download and the prompts you send to your own server
• Versions: others have published 7 quantized and 1 fine-tuned copy

On Hugging Face, 13 November 2026

• Parameters: 19.9B
• Weights: about 19.9 GB in F8_E4M3; about 9.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 1,939,275
• Likes: 540
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 7
• Fine-tunes: 1
• Tags: safetensors, qwen3_5, unsloth, compressed-tensors

Links

• Model: https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4
• Files: https://huggingface.co/unsloth/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 unsloth/Qwen3.8-27B-NVFP4

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