gpt-oss-safeguard-120b

gpt-oss-safeguard-120b is part of OpenAI's gpt-oss family and is purpose-built for safety classification rather than general chat. Released as a research preview, it is a fine-tuned descendant of OpenAI's earlier gpt-oss open models and is offered under the same permissive Apache 2.0 license, making the full weights freely available for use, modification, and deployment. Its design treats safety as an in-context reasoning problem: developers supply their own policy at inference time, and the model classifies user messages, completions, or full conversations against that policy, rather than relying on a decision boundary baked in during training.

The practical appeal of this approach is flexibility and transparency. Because the policy lives in the prompt, teams can iteratively refine their safety rules without retraining, adapting the classifier to new products, regions, or risk profiles. The model exposes its chain-of-thought so reviewers can audit how a given classification was reached, which helps with debugging false positives and documenting policy enforcement. With 120 billion parameters, it sits at the larger end of the open-weight safety-classification options and is well suited for organizations that want customizable, inspectable content moderation grounded in their own definitions of acceptable and unsafe behavior.