Qwen3.8-27B

Qwen3.8-27B

Reads text, images and video and replies in text, with a 27B open model tuned for coding, research and multi-step agent work. Suits developers who run models themselves.

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: Qwen3.8-27B is an open model from Qwen that reads text, images and video and replies in text. It is a dense model with 27 billion parameters, aimed at coding, research, professional work and agent tasks with many steps. It belongs to the Qwen3.8 generation, which builds on the Qwen3.5 design.

Who it suits: It suits developers and researchers who want to run an open model on their own servers, and teams building agents that must read screenshots, documents or video.

Industry | Who uses it | How they use it | Example
Software | Developers | Run it as the coding model behind their own agent framework or dev tools | Ask it to find why a test fails and propose a fix
Research | Analysts and scientists | Give it papers, charts and diagrams and ask for explanations | Send a STEM diagram and ask what it shows
Agents | Teams building automations | Let it plan a multi-step task and react to what the tools return | Give it a job that needs several tool calls and let it carry on
Media | Reviewers and editors | Ask questions about a long recording | Send an hour-long video and ask where a topic comes up

If you only want a hosted service with no servers to look after, you can wait, because the model card says a hosted version is coming soon.

Where to use it:

• Hugging Face Transformers: the weights are published in its format
• vLLM: listed as compatible
• SGLang: listed as compatible
• TokenSpeed: listed as compatible
• Coding frameworks and developer tools: the card says support for popular ones is broader than before

Use case | Who | What they get back
Coding help | Developers | Code written or changed across a task with several steps
Reading documents and diagrams | Researchers | A written explanation of an image or page
Questions about video | Analysts | Answers about what happens in a long video

How to set it up: You need these first.

• One of the runtimes the page names: Transformers, vLLM, SGLang or TokenSpeed
• A machine with enough memory for a 27B model, since the page gives no exact figure

1. Choose a runtime from that list and install it by its own guide.
2. Download the Qwen3.8-27B weights from the model page.
3. Load the model in your runtime from the downloaded files.
4. Send a test prompt with an image and check that a text answer comes back.

The card text available here shows no install command, so follow the guide of the runtime you chose.

How to use it: You send text, images or video to the model through your runtime and read the text that comes back. Thinking mode is on by default. You can turn it off for one request and set how deep the reasoning goes.

Tool | What it does
Thinking mode | On by default and can be switched off for a single request
reasoning_effort | Sets how deep the model reasons
preserve_thinking | Keeps the reasoning from earlier messages in the conversation

Example: You attach a screenshot of a chart and ask what trend it shows and what looks odd. The model reads the image, reasons first if thinking is on, and answers in text.

For newcomers and experts: It is a chat model that you run yourself, and it can look at pictures and video as well as read text. The first small thing to try is one image with one short question, with thinking left on.

Under the hood it is a dense model, so all 27B parameters take part in every token. It has 64 layers, built as 16 repeats of three Gated DeltaNet layers followed by one Gated Attention layer, each with a feed-forward network. A vision encoder handles images and video, and training also taught it to predict several tokens ahead. Context is 262,144 tokens natively and can be extended to 1,000,000.

Text, image or video goes in
Vision encoder reads images and video
Layers alternate Gated DeltaNet and Gated Attention
A thinking step runs unless switched off
A text answer comes back

Good to know:

• License: Apache-2.0
• The context window is 262,144 tokens natively and can be extended to 1,000,000
• The weights use the Hugging Face Transformers format, and the card lists Transformers, vLLM, SGLang and TokenSpeed as compatible
• Qwen Cloud is announced as a hosted option with 1M context by default and built-in tools, marked as coming soon
• The card text available here states no hardware requirement

One benchmark row in the card compares it with the earlier Qwen3.6-27B.

Title: Terminal Bench 2.1 agentic terminal coding score
Qwen3.8-27B: 73.0 points
Qwen3.6-27B: 63.4 points

On Hugging Face, 10 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,508,158
• Likes: 19,504
• License: apache-2.0
• Quantized versions: 1,730
• Fine-tunes: 858
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, conversational

Links

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

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