Qwen publishes residual-stream SAE for Qwen3.5-9B-Base (64K width, L0=50)
Qwen published a TopK sparse autoencoder trained on the residual stream of Qwen3.5-9B-Base, with 64K width and tighter L0 sparsity of 50, for mechanistic interpretability research. Published on Hugging Face under the Qwen organisation.
PUBLISHED2026-04-27
OBSERVED2026-08-11
AGE4mo
SOURCES1
- Repository:
Qwen/SAE-Res-Qwen3.5-9B-Base-W64K-L0_50 - Tags: topk_sae, sparse-autoencoder, sae, mechanistic-interpretability, qwen-scope
- Associated paper: arxiv:2605.11887
- License: other (not specified)
COMMUNITY
No curated reactions recorded for this event. Facts and takes are kept in separate layers — community context is added by hand, never blended into the record above.