GPT OSS Safeguard 20B

On October 29, 2025, OpenAI released gpt-oss-safeguard as a research preview of open-weight safety reasoning models in two sizes, 120b and 20b, available under the Apache 2.0 license on Hugging Face. The models interpret a developer-provided policy at inference time to classify user messages, completions, and chats, using chain-of-thought that developers can review. They are fine-tuned versions of the gpt-oss open models. Unlike traditional classifiers that infer decision boundaries from labeled examples, gpt-oss-safeguard takes both a policy and content as inputs and outputs a classification with reasoning, enabling rapid policy iteration. The 20b variant is positioned alongside the 120b sibling for safety classification, content labeling, and trust-and-safety pipelines where developers control policy definitions. OpenAI recommends its original gpt-oss models for general end-user applications.

The DeepWiki technical documentation for gpt-oss-safeguard-20b, last indexed October 30, 2025, confirms the variant's sparse-activation Mixture-of-Experts architecture with 21 billion total parameters and 3.6 billion active parameters, a 17% activation ratio. The model requires the harmony response format and is optimized for lower-latency, resource-constrained safety classification deployments. The page covers deployment topics including hardware requirements, model download and setup, inference configuration, policy file structure, severity levels, dual-label depiction-versus-request handling, and golden dataset validation. It is positioned as the smaller sibling to gpt-oss-safeguard-120b within the broader gpt-oss-safeguard family and is released under the Apache 2.0 license.

OpenAI's technical report landing page for gpt-oss-safeguard reiterates that both the 120b and 20b variants are open-weight reasoning models post-trained from gpt-oss, available under the Apache 2.0 license and the gpt-oss usage policy. The models are customizable, expose the full chain-of-thought, support reasoning effort levels of low, medium, and high, and are compatible with Structured Outputs and the Responses API. The page confirms the safeguard models are fine-tunes of their gpt-oss counterparts and were trained without additional biological or cybersecurity data, extending prior worst-case safety assessments. OpenAI recommends classifying content against a provided policy with these models and using the original gpt-oss line for end-user chat, noting the safeguard models are not intended for direct user interaction.

OpenAI's technical report for gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, dated October 29, 2025, describes the models as open-weight reasoning checkpoints post-trained from gpt-oss to reason from a provided policy and label content accordingly. The text-only models support configurable reasoning effort (low, medium, high), expose the full chain-of-thought, and are compatible with the Responses API and Structured Outputs. The report outlines sections on safety classification performance, multilingual performance, and observed challenges including disallowed content, jailbreaks, instruction hierarchy, hallucinated chains of thought, hallucinations, and fairness and bias. OpenAI notes the safeguard models were trained without additional biological or cybersecurity data, so prior worst-case estimates from the gpt-oss release carry over, and recommends using the original gpt-oss models for general end-user chat applications.

The official OpenAI Hugging Face model card for gpt-oss-safeguard-20b specifies a 21B-parameter Mixture-of-Experts design with 3.6B active parameters, fitting into GPUs with 16GB of VRAM. The model requires the harmony response format and is intended for safety classification tasks, with configurable reasoning effort set to low, medium, or high depending on latency needs. It is released under the permissive Apache 2.0 license. Key highlights include training for safety reasoning, a bring-your-own-policy design that interprets written policies, and full access to the model's chain-of-thought for debugging. gpt-oss-safeguard is a model partner of the ROOST Model Community, a group of safety practitioners exploring open-source AI models for online protection. The model card notes raw chain-of-thought is meant for developers and safety practitioners, not general users.