release aidev NEEDS REVIEW SIG 1/5

Qwen publishes SAE residual stream weights for Qwen3.5-2B-Base (W32K, L0 50)

Qwen released a sparse autoencoder (SAE) residual stream decomposition for the Qwen3.5-2B-Base model, configured with width 32K and L0 sparsity target 50. Part of the Qwen-Scope mechanistic interpretability research effort (arXiv:2605.11887). No capability or benchmark data is provided in the source.

PUBLISHED2026-04-27
OBSERVED2026-08-11
AGE4mo
SOURCES1

Model repository Qwen/SAE-Res-Qwen3.5-2B-Base-W32K-L0_50 published on Hugging Face.

  • Tags: topk_sae, sparse-autoencoder, sae, mechanistic-interpretability, interpretability, qwen-scope, arxiv:2605.11887
  • Base model: Qwen/Qwen3.5-2B-Base
  • License: other (not specified)
featuremodel-supportcapability

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.