An abliterated model is an open-weight 📝Large Language Model (LLM) whose built-in tendency to refuse requests has been removed by editing its weights directly, without retraining, so it answers prompts its original safety tuning would have declined.
The technique rests on a June 2024 paper by Andy Arditi, Neel Nanda, and colleagues, "Refusal in Language Models Is Mediated by a Single Direction," which found across 13 open chat models of up to 72 billion parameters that refusal is controlled by one direction in the model's internal activations. Removing that direction stops the model from refusing harmful instructions, and adding it makes the model refuse even harmless ones. Abliteration, a blend of "ablation" and "obliterate," estimates the direction by comparing activations on harmful and harmless prompts, then subtracts it out of the weight matrices that write to the model's residual stream.
Because it is a one-time weight edit, abliteration is cheap and leaves no runtime trace, and abliterated variants of popular open-weight families circulate widely on 📝Hugging Face, usually with "abliterated" in the name. The method is only possible where weights are public, which makes it one of the clearer practical stakes in 📝Open-Weight vs. Closed-Weight AI Models: safety behavior trained into released weights can be stripped out by anyone. Abliteration also tends to cost some general capability, with reported effects on math and reasoning benchmarks ranging from negligible to substantial depending on the model and tool, and newer model generations spread refusal across more layers, making a single-direction edit less complete.