OpenAI releases open safety reasoning models
OpenAI has introduced two new open-weight models, called gpt-oss-safeguard (20b and 120b). Both models are designed for companies that prioritize safety.
The details
-
The new gpt-oss-safeguard models come in two sizes (20b and 120b parameters). Both models are fine-tuned versions of the gpt-oss open models. The new models are available under the same permissive Apache 2.0 license as gpt-oss, which allows developers to use, modify, and deploy them freely.
-
In addition, the new models use reasoning to directly interpret a policy provided by the developer at inference. The developer can always decide which policy to use, making the responses more relevant.
-
According to OpenAI, the models use a chain-of-thought process that developers can review to understand how the model makes its decisions.
-
OpenAI collaborated with ROOST to develop gpt-oss-safeguard. ROOST (Robust Open Online Safety Tools) is a nonprofit open-source initiative that develops and distributes accessible AI safety tools to help organizations build scalable, resilient, and responsible safety infrastructure.
-
The gpt-oss-safeguard models are available via Hugging Face and Ollama.
Our thoughts
We have already tested gpt-oss in our step-by-step guides and think it is a great model for the open-source community. In addition, you can use gpt-oss-safeguard freely without copyleft restrictions or patent risk. That makes it ideal for experimentation, customization, and commercial deployment.
More information: 🔗 OpenAI
Magic AI tools
Boost Your Productivity with AI
Explore the best AI tools to boost your efficiency and productivity.
