On April 30, 2026, Ali QianWen made a significant announcement regarding the open-source availability of Qwen-Scope, an innovative interpretability module. This module has been meticulously trained using the Qwen3 and Qwen3.5 series models as its foundation. By incorporating Sparse Autoencoders (SAE) into the hidden layers, Qwen-Scope is able to extract features that are not only highly decoupled and exhibit low redundancy but are also more amenable to interpretation. The open-source release encompasses weights for seven large-scale models, spanning both dense models and mixture-of-experts models, with a comprehensive collection of 14 sets of SAE weights. Qwen-Scope is designed to cater to a wide array of application scenarios, including inference-directed control, data classification and synthesis, identification and optimization of training anomalies, as well as redundancy analysis in evaluation sets. This groundbreaking module has been made available on the renowned Hugging Face and ModelScope platforms, inviting developers and researchers to explore its full potential.