Ollama has announced a partnership with OpenAI and ROOST to distribute the gpt-oss-safeguard reasoning models. These models are designed specifically for safety classification tasks, including LLM input-output filtering and content labeling.
The models are available in two sizes: 20B and 120B. They operate under a permissive Apache 2.0 license, allowing free experimentation, customization, and commercial deployment without copyleft restrictions or patent risk.
Reasoning and Policy Handling
The models are trained to reason about safety, interpreting written policies to generalize across products with minimal engineering. Unlike standard classifiers, they provide access to the reasoning process rather than just a score.
OpenAI evaluated these models on internal sets where multiple policies were provided at inference time. Accuracy is determined by exact matching against golden set labels for all included policies. External evaluations included OpenAI's 2022 moderation dataset and the ToxicChat public benchmark.
Developer and Safety Use Cases
Vinay Rao, CTO of ROOST, noted that gpt-oss-safeguard is the first open-source reasoning model with a 'bring your own policies' design. The models demonstrate skill in understanding different policies and explaining their reasoning.
Reasoned decisions are intended for developers and safety practitioners. Raw Chain of Thought (CoT) output is not meant for exposure to general users or use cases outside of safety contexts.
Users can configure the reasoning effort as low, medium, or high based on specific latency needs.
