Qwen releases sparse autoencoders for its models
Alibaba's Qwen team has open-sourced Qwen-Scope, 14 sets of sparse autoencoders trained on seven Qwen3 and Qwen3.5 models, giving researchers a feature-level view into the internals of a major open-weight model family.
The Qwen team introduced Qwen-Scope, an interpretability toolkit built from sparse autoencoders (SAEs) inserted into the hidden layers of Qwen models and trained there. The SAEs apply sparsity constraints to break the models' dense hidden representations into sparse features that people can inspect. The weights are on Hugging Face and ModelScope, and a technical report accompanies them.
The release covers seven models, both dense and mixture-of-experts. The dense models are Qwen3-1.7B, Qwen3-8B, Qwen3.5-2B, Qwen3.5-9B and an instruct version of Qwen3.5-27B. The MoE models are Qwen3-30B-A3B and Qwen3.5-35B-A3B. Each model gets two SAEs, one with an L0 sparsity of 50 and one with 100. Widths run from 32K to 128K features, and expansion factors are 16 or 64. All of the SAEs were trained on 0.5B tokens sampled from the matching model's pretraining data.
Qwen describes four uses:
- Inference: changing feature activations to steer language, entities or style without written instructions.
- Data: classifying toxic text from a small set of seed examples with no extra training. Qwen also uses the features to generate examples for toxic-text features that rarely or never fire, which it says makes training data about 15 times more efficient.
- Training: a fine-tuning loss aimed at activations linked to unwanted language mixing, and amplifying features linked to repetitive output so that reinforcement learning samples those failures more often.
- Evaluation: finding benchmarks whose activated features overlap, to show which test sets are redundant.
The blog post does not give detailed benchmark results for these claims. It points readers to the technical report for more detail.