Abliterated Model is an open-weight large language model distributed by abliteration.ai that has undergone abliteration, a post-training technique which identifies the model's learned refusal direction in activation space and dampens it so that the model is less likely to refuse a prompt. The technique, originally implemented in FailSpy's abliterator library, surgically removes the refusal direction from selected representations rather than relying on prompt-level jailbreaks, leaving the model's underlying language, reasoning, and tool-use capability intact. In practice this means the model behaves as a less-censored counterpart to a provider-hosted system, which is useful when refusal defaults get in the way of authorized research, synthetic data generation, red teaming, or evaluation work where application-layer policy controls are owned by the developer.

According to abliteration.ai's model documentation, this release is positioned as the general-purpose default in the provider's catalog and accepts image inputs in addition to text, while a larger sibling model in the same family is text-only and optimized for longer reasoning workloads. The documentation lists streaming, tool and function calling, structured output modes, and controllable reasoning effort as supported capabilities, with reasoning traces available in responses and the option to disable or hide them. Because abliteration is a permanent weights-level modification rather than a session-level prompt trick, teams can deploy the model with their own policy and governance layer on top, using it for security research, model behavior analysis, and governed public-sector workflows where the default refusals of hosted models would otherwise interfere with the task.