QwenScope

QwenScope

Explore and steer the Qwen3.5-2B language model through sparse autoencoder features: compare two texts' feature activations, or push a feature into generation. For interpretability researchers.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: QwenScope is a Gradio app for looking inside the Qwen3.5-2B language model with a sparse autoencoder, or SAE. An SAE breaks the model's internal activity into many separate features that you can read and test.

The app has two parts. One compares which features fire on two pieces of text, and the other adds a feature to the model's work to steer what it writes.

Who it suits: It suits researchers and students of model interpretability who want hands-on access to SAE features. Developers who study how steering is wired can read app.py too.

Industry | Who uses it | How they use it | Example
AI research | Interpretability researchers | Compare feature activations between two texts | Checking which features differ between a polite and a rude sentence
AI safety | Safety researchers | Steer generation with a chosen feature | Seeing how one feature changes the model's output
Education | Students | Try both tools on short texts | Learning what a sparse autoencoder feature is
AI development | Developers | Read app.py and requirements.txt | Reusing the idea for another model

Skip it if you want a normal chat with the model, since this app is for studying the model's internals.

Where to use it: The app runs in these places.

• A web browser, on the Space's page at Hugging Face
• A local copy on Windows, macOS or Linux, running Python 3.12

Use case | Who | What they get back
Compare two texts | Researchers | Feature activations across all transformer layers
Steer generation | Researchers | Model output changed by an injected feature
Study the demo | Developers | A short Gradio app around one model and one SAE

How to set it up: Online you only need a browser. A local copy needs these.

• Python 3.12 on Windows, macOS or Linux
• The pip command
• Room on disk for the Qwen3.5-2B model and its SAE

1. Install the huggingface_hub package.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Install gradio at version 5.49.1.
5. Install the packages listed in requirements.txt.
6. Start the app with python app.py.

pip install -U huggingface_hub
hf download Qwen/QwenScope --repo-type space --local-dir QwenScope
cd QwenScope
pip install gradio==5.49.1
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page, or run the command below in your copy and open the page the app serves. Then pick one of the two tools and give it text.

python app.py

Tool | What it does
Feature Comparison | Shows how strongly SAE features respond to two texts, layer by layer through the model
Feature Steering | Mixes chosen decoder vectors into the model's residual stream and changes what it writes

Example: You open Feature Comparison and enter two short texts that differ in tone. The app shows which SAE features activate differently in each layer. Then you try Feature Steering with one of those features.

For newcomers and experts: If you are new, open the Space's page and start with Feature Comparison on two short, similar sentences. Differences are easier to spot when little changes between the texts.

If you know the field, the SAE has a width of 32,768 features and a Top-K of 100. It covers all 24 layers of a model with a hidden size of 2,048. Steering adds decoder vectors back into the residual stream. The README credits Hugging Face ZeroGPU for the compute.

Open the app and pick a tool
Enter one or two texts
The SAE reads the model's activations
You compare features or steer output

Good to know: The README gives the model setup in a short table.

Property | Value
Base model | Qwen3.5-2B
SAE width | 32,768
Top-K | 100
Layers | 24 in all, numbered 0 to 23
Hidden size | 2,048

• License: apache-2
• Cost to run: none named. The README says the Space uses Hugging Face ZeroGPU
• Data it reaches: only the Qwen model and SAE repositories it names. The source says nothing else.

On Hugging Face, 13 November 2026

• Likes: 45
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/Qwen/QwenScope
• Files: https://huggingface.co/spaces/Qwen/QwenScope/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.