Qwen3.8-27B-FP8

Qwen3.8-27B-FP8

Handles text, images and video with FP8-quantized weights of Qwen3.8-27B, a 27B dense model with adjustable thinking. For developers serving it with transformers, vLLM or SGLang.

Made by Qwen 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 repository holds FP8-quantized weights of Qwen3.8-27B in Hugging Face Transformers format. Qwen3.8-27B is a dense vision-language model that reads images and video, and it is built for coding, research and long multi-step agent tasks. The card says the FP8 version performs nearly the same as the original.

Who it suits: It suits developers who want a compact dense model for coding and agent work with less memory than the full-precision files need, and teams who serve models with vLLM or SGLang.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a 27B model for coding and agent tasks | Ask for help on a multi-step coding job
Research | Analysts | Use thinking control for deeper reasoning on hard questions | Raise reasoning effort for a research question
Hosting | Teams serving models | Load the FP8 files in vLLM or SGLang | Serve a vision-language model to internal tools

People who want a managed service without running anything can skip the weights, since the card mentions an official hosted option from Qwen Cloud.

Where to use it:

• Hugging Face Transformers, in the format the repository holds
• vLLM, SGLang and TokenSpeed, which the card says the files work with
• Windows, macOS or Linux, where you can download the files with Python

Use case | Who | What they get back
Agentic coding | Developers | Planning and better handling of environment feedback across long tasks
Image and video understanding | Analysts | Answers about diagrams, documents and long videos
Reasoning control | Application builders | Thinking that is on by default and can be turned off per request

How to set it up: You need these first:

• Python and the huggingface_hub package
• Room for the files, about 27.8 GB as F8_E4M3 or about 13.9 GB at 4-bit as an estimate
• One of the engines the card names, such as Transformers, vLLM or SGLang

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load them in Transformers, vLLM, SGLang or TokenSpeed.

pip install -U huggingface_hub
hf download Qwen/Qwen3.8-27B-FP8

How to use it: Download the files, load them in an engine the card names, then chat with the model, with images or video if you like. The card text here has no run commands, so follow its model card for the exact calls.

Tool | What it does
Thinking mode | On by default and can be disabled for each request
reasoning_effort | Tunes how deeply the model reasons
preserve_thinking | Keeps reasoning context from earlier messages

Example: A developer could ask the model to plan and carry out a multi-step coding task, with reasoning_effort set higher for a hard part. The card says the model is tuned to carry such tasks through to completion.

For newcomers and experts: This is a mid-sized model that reads text, pictures and video. It is quantized to FP8, which makes the files smaller. A first small step is to download them and read the card's notes on thinking control.

The card says the quantization is fine-grained FP8 with a block size of 128. The model has 64 layers laid out as sixteen repeats of three Gated DeltaNet blocks and one Gated Attention block, each followed by a feed-forward network. The hidden dimension is 5120, and multi-token prediction was trained with several steps. Context is 262,144 tokens natively and can reach 1,000,000.

Download the FP8 weights with hf
Load them in a supported engine
Send text, images or video
Set thinking and effort as needed
Read the reply

Good to know:

• Size: 27.8B parameters per Hugging Face, about 27.8 GB in F8_E4M3 and about 13.9 GB at 4-bit, as an estimate
• Versions: Hugging Face lists 9 quantized and 8 fine-tuned versions made by other people
• Hosted option: the card says an official API from Qwen Cloud is coming, with more production features such as 1M context by default

Title: Context length in tokens
Native: 262,144 tokens
Extended: 1,000,000 tokens

• License: apache-2.0
• Cost to run: none named for the local files, no API key or account. The hosted Qwen Cloud service is separate
• Data it reaches: it downloads model files from Hugging Face, and the card text names no other service for the local run

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 27.8 GB in F8_E4M3; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 5,442,810
• Likes: 984
• License: apache-2.0
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 9
• Fine-tunes: 8
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational, fp8

Links

• Model: https://huggingface.co/Qwen/Qwen3.8-27B-FP8
• Files: https://huggingface.co/Qwen/Qwen3.8-27B-FP8/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen3.8-27B-FP8

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