An uncensored edit of Qwen/Qwen3-8B. Refusal behavior is removed by editing the weights directly — no finetuning, no adapter, no runtime hook. The result is a standard Transformers checkpoint that loads anywhere Qwen3 does.

Produced with Apostate.

Apostate identifies the residual-stream direction that separates refused prompts from answered ones, then projects it out of the model's weights permanently. For Qwen3-8B (a standard pre-norm dense transformer), the edit targets the writer side: the refusal direction is removed from the weight matrices of every module that writes to the residual stream — attention output projections and MLP down-projections — across all layers.

The edit uses oblique (mean-preserving) ablation: the operator E = I − R Uᵀ where U is R minus its harmless-mean component. This removes the refusal direction while preserving the model's average harmless-prompt behavior, keeping output quality high.

The refusal subspace is found with a rank-3 predictive TPE search, with causal layer importance scoring to focus edits on the layers that most drive refusal (concentrated in the mid-to-late layers, peak at layer 29).

Evaluated on held-out prompts from JailbreakBench and the harmful_behaviors test split. Refusal is graded by a classifier with a weak-compliance guard; KL measures token-distribution shift on harmless prompts.

The model answers freely on requests the base model refuses while remaining coherent and on-task for everyday use.

Qwen3 supports a thinking mode — pass enable_thinking=True to the chat template if you want extended reasoning.