Qwen-Scope

#42 in Reasoning Models

alibaba · since 2026-04-29 · 2× · last seen Jun 30, 2026

3

Momentum

Qwen-Scope is an open-source interpretability toolkit released by Alibaba's Qwen team, consisting of a suite of Sparse Autoencoders (SAEs) trained within the hidden layers of Qwen3 and Qwen3.5 base models. It decomposes dense model activations into sparse, disentangled, and interpretable features. Practical use cases include inference-time steering, data classification and synthesis, evaluation analysis, and targeted post-training. It is not a standalone language model and has no API pricing, context window, or reasoning benchmarks (e.g., AIME).

Sources (2)