GPT OSS Safeguard 120B

OpenAI released gpt-oss-safeguard on October 29, 2025 as a research preview of open-weight reasoning models purpose-built for safety classification, available in 120b and 20b sizes under Apache 2.0. Both variants are fine-tunes of the existing gpt-oss open models and can be downloaded today from Hugging Face. Unlike conventional classifiers that bake policy into weights, gpt-oss-safeguard reads a developer-supplied policy at inference time and classifies user messages, completions, or full chats against it, returning its chain-of-thought reasoning. This lets developers iterate on policies quickly for evolving harms, nuanced domains, or limited labeled data scenarios.

Ollama announced (October 29, 2025) a partnership with OpenAI and ROOST to host gpt-oss-safeguard 20B and 120B for safety classification tasks, distributed under Apache 2.0. Users can pull the models directly via Ollama after installing the runtime. The integration highlights training for safety reasoning, bring-your-own-policy interpretation, full chain-of-thought visibility for developers, and configurable reasoning effort (low, medium, high). Raw CoT is explicitly scoped to developers and safety practitioners, not general end users.

OpenAI's gpt-oss-safeguard technical report (October 29, 2025) details the 120b and 20b variants as text-only open-weight reasoning models post-trained from gpt-oss, released under Apache 2.0 and compatible with OpenAI's Responses API. The models expose full chain-of-thought, configurable reasoning effort (low, medium, high), and Structured Outputs support. The report recommends using gpt-oss-safeguard to classify content against a provided policy rather than as the end-user-facing model, and shares baseline safety evaluations across disallowed content, jailbreaks, instruction hierarchy, and multilingual settings. It also notes no additional biological or cybersecurity training data was used, so prior gpt-oss worst-case estimates apply.