Qwen-Scope

Qwen-Scope is an open-source interpretability release from the Qwen Team: a suite of sparse autoencoder (SAE) weights/modules trained on Qwen3 and Qwen3.5 large language models. Official sources describe 14 groups of SAEs across 7 Qwen backbones, used for language-model interpretability, inference-time steering, evaluation analysis, data workflows, and post-training optimization.

Model Intelligence