A custom abliteration run on Alibaba’s Qwen 3.8 27B removed the model’s hard refusal to generate political satire about China, allowing it to answer prompts it had previously blocked.
The method targeted a “refusal vector” rather than retraining the model: activations from refusal and harmless prompts were compared across 64 layers, with the strongest contrast found at layer 60.
The developer’s five-stage pipeline extracted the roughly 56 GB checkpoint, tested candidate edits in memory, then baked the winning projector into residual-writing modules across layers 1–63 on an NVIDIA GB10 with 128 GB memory.
The full process took about an hour, used deterministic settings with temperature and thinking both set to 0, and relied on two related prompts after a single-prompt attempt failed to isolate the censorship behavior.
The result, published to HuggingFace for review, underscores a broader claim that open-weight models can carry embedded state or corporate controls even when they run locally.