OpenAI Releases New Open-Weight Reasoning Model

Actually, there are even two gpt-oss-safeguard models—in sizes 120b and 20b. Both are open-weight reasoning models from OpenAI and are intended for use in companies. The focus during development was on the possible implementation of security measures—hence the name. Companies have very different requirements for what a model can do or where the limits of the applications built on it lie.

In addition, the models can be adapted to one's own infrastructure. They are available under the Apache 2.0 license. They are available on Hugging Face. As is usual with AI, this is a preview.

That the models are specialized in security measures means that they can follow guidelines particularly well. For this, the respective model uses reasoning, i.e., the reasoning capability. Guidelines that developers can set themselves are interpreted at the time of inference, according to OpenAI. Furthermore, the model's chains of thought can be viewed and checked. This is intended to help understand how models arrived at decisions. Based on this, further measures can then be taken. Since the models process the specified guidelines when processing a request, it is possible to iteratively define further guidelines when the need arises. Other models, as OpenAI writes, that learned guidelines during training could not react as well to new requirements.

Since the OpenAI blog post also states that the new models are used when latency is not as important as high-quality responses, the reasoning process is likely to slow down the response somewhat.

In the development of gpt-oss-safeguard, OpenAI collaborated with ROOST. This is a community dedicated to building a robust, open, and secure AI infrastructure.

Security Measures for Closed Models

Those who use OpenAI's closed models can, for example, access the real-time API with an integrated moderation system. Developers can cancel the output here as soon as the conversation goes in an undesirable direction. Of course, all GPT models come with built-in guidelines and guardrails.

Most recently, OpenAI introduced further security measures for ChatGPT. This concerns the protection of individuals with mental health problems.

(emw)