Qwen3-30B-A3B MoE-Output Sparse Autoencoders

About

TopK sparse autoencoders (k = 100, 2^20 latents) trained on the MoE block output of Qwen3-30B-A3B at layers 4, 30 and 46. They were released with a paper reporting that MoE experts show human-interpretable, fine-grained specialization, so they let researchers inspect which features the experts' outputs encode. Each SAE file is about 17.2 GB in float32, and the model card gives a loading and encoding code example.