Built upon the gpt-oss-20b architecture, this model utilizes a 21-billion parameter Mixture-of-Experts design with 3.6 billion active parameters. It is specifically engineered to excel at safety-focused reasoning tasks, such as filtering LLM inputs and outputs, labeling online content, and performing offline safety assessments. By providing a dedicated tool for evaluating text against custom safety policies, it serves as a specialized utility for developers who need to maintain rigorous content standards within their applications.

The model is trained using the harmony response format, which is a requirement for its proper operation and integration. Its design lineage emphasizes a balance between performance and latency, making it a practical choice for high-throughput safety filtering. Because it is released under the permissive Apache 2.0 license, it offers flexibility for developers to modify and deploy the model within their own infrastructure, supporting a wide range of safety-critical workflows that require reliable and consistent content analysis.