---
title: QwenScope
emoji: 🔥
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
python_version: '3.12'
app_file: app.py
pinned: false
license: apache-2.0
---

# QwenScope Feature Explorer

Interactive Sparse Autoencoder (SAE) feature exploration and steering demo for Qwen3.5-2B, powered by Hugging Face ZeroGPU.

## Features

- **Feature Comparison**: Compare SAE feature activations between two text examples across all transformer layers.
- **Feature Steering**: Inject SAE decoder vectors into the residual stream to steer model generation behavior.

## Model Info

| Property | Value |
|---|---|
| Base model | [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B) |
| SAE width | 32,768 |
| Top-K | 100 |
| Layers | 0 – 23 (24 total) |
| Hidden size | 2,048 |

## Citation

```bibtex
@misc{qwenscope,
  title  = {{Qwen-Scope}: Turning Sparse Features into Development Tools for Large Language Models},
  author = {{Qwen Team}},
  month  = {April},
  year   = {2026}
}
```