Abliteration computes a "refusal direction" by averaging the activation difference between harmful and harmless prompts across a model's layers, then orthogonally projects that direction out of the weights so it no longer influences outputs. Unlike fine-tuning, it needs no training data or GPU training run — it is an inference-time computation applied once, permanently, to the model's weights. Tools like Heretic automate the process; commercial vendors including OrcaRouter and Abliteration.ai have applied it to models like Qwen3.8-27B and GLM-5.3, either as downloadable weights or as a hosted API marketed to red teams and cybersecurity testers.