OpenAI has proudly declared the release of two innovative open-weight models dedicated to enhancing AI safety: gpt-oss-safeguard-120b and gpt-oss-safeguard-20b. These models represent refined iterations of the previously introduced gpt-oss series open models and are distributed under the Apache 2.0 license, granting users the freedom to utilize, adapt, and deploy them as needed. The gpt-oss-safeguard-120b model boasts an impressive total of 117 billion parameters, with 5.1 billion parameters activated for each token. It is ideally suited for production-level, general-purpose applications that demand high reasoning capabilities and can operate efficiently on a single Nvidia H100 GPU equipped with 80GB of memory. In contrast, the gpt-oss-safeguard-20b model features 21 billion parameters and can function seamlessly on a consumer-grade laptop with just 16GB of memory, making it an excellent choice for local reasoning and device-end applications. Both models accommodate context lengths extending up to 128k tokens, leveraging a combination of alternating dense and locally banded sparse attention mechanisms. They also employ grouped multi-query attention techniques to boost reasoning efficiency. Furthermore, these models incorporate advanced features such as chain-of-thought reasoning, tool calling, and structured output, rendering them exceptionally well-suited for constructing agent workflows.