Qwen-Scope: Official Sparse Autoencoders (SAEs) for Qwen 3.5 models
Qwen Team released Qwen-Scope — a collection of Sparse Autoencoders (SAEs) for the Qwen 3.5 family (from 2B to 35B MoE). They’ve mapped internal features for the residual stream across all layers. What is this exactly? Think of it as a dictionary…