Cross-Layer TopK CLT for Qwen3-8B-Base

About

A cross-layer transcoder for Qwen3-8B-Base with 36 layers and 32,768 latents per layer. Each layer's latents read the residual stream before that layer's MLP and write to the MLP outputs of that layer and every later layer, with the 360 highest-scoring latents kept jointly across all layers. It supports mechanistic interpretability work on how the base model's MLP computation is spread across layers. The checkpoint is 94.2B parameters in fp32 (377 GB) and the model card includes a multi-GPU encoding demo.