As an open-weight safety reasoning model in the gpt-oss family, this release is positioned as a fine-tuned descendant of OpenAI's earlier gpt-oss open models, distributed under the same permissive Apache 2.0 license. A 120B-parameter variant (reported as roughly 117B total with around 5.1B active parameters in a mixture-of-experts configuration) is paired with a smaller 20B sibling, and both are available for download from the OpenAI organization on Hugging Face. Hosting providers such as Fireworks AI additionally expose the weights for on-demand deployment, broadening access beyond direct self-hosting.

The defining design choice is that the model is built to interpret a developer-supplied safety policy directly at inference time rather than relying on a fixed, trained-in classifier. It can label user messages, completions, or full chats, and it returns chain-of-thought reasoning that developers can audit to understand each decision. Because the policy is injected per request, teams can iterate on rule wording, expand coverage, or swap policies for new product contexts without retraining, which makes the model a flexible fit for organizations that want transparent, customizable content moderation rather than a black-box safety filter.