Qwen-Scope (sparse autoencoders for Qwen3 and Qwen3.5)
Open suite of sparse autoencoders for Qwen3 and Qwen3.5 (14 SAE groups, 7 model variants), used for steering, evaluation analysis and data work.
Covers dense and MoE models. The paper shows SAE features steering language, concepts and preferences at inference without touching weights, probing benchmark redundancy and guiding data decisions, in addition to post-hoc analysis.
- Date
- Wednesday, 29 April 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SAE groups / model variants | 14 / 7 Qwen3 and Qwen3.5, dense and MoE | company |
Blog dated 2026-04-29 per Releasebot; SAE repos created on Hugging Face 2026-04-27; paper arXiv 2605.11887 submitted 2026-05-12. Claims are the authors' own.
Sources
- releasebot.io/updates/qwen
- huggingface.co/api/models?author=Qwen&sort=createdAt&direction=-1&limit=60
- arxiv.org/abs/2605.11887
This record was checked against its sources on 6 October 2026. How we check