OpenAI has announced new models called gpt-oss-120b and gpt-oss-20b, which are open-weight, meaning their internal parameters, or weights, are freely shared for public use. These artificial intelligence (AI) models were trained with special steps to avoid harm. During pre-training, harmful information about dangerous topics like chemicals, biology, radiation, and nuclear issues was removed. After that, a process called deliberative alignment taught the models to say no to unsafe requests and resist tricks that could make them act badly. This helps keep the models secure.
To test risks, the models were fine-tuned, which means adjusted with specific data, to see if they could be used for bad purposes, like in biology or cybersecurity. These steps show progress in making open models safer, leading to the decision to release them.
Availability and use
The gpt-oss models are free to download on Hugging Face, a platform for sharing AI tools, and work with low memory needs - 80GB for gpt-oss-120b and 16GB for gpt-oss-20b. Example tools and support for platforms like PyTorch and Apple’s Metal are also provided. Partnerships with companies like Azure and NVIDIA ensure the models run well on different systems, including Windows devices with Microsoft’s help.
These models are flexible for developers to adjust and use locally or online. For those needing more features, the API platform is better, though API support for gpt-oss might come later based on feedback. A playground and guides are available to try and learn about the models.
Open-weight models like gpt-oss bring new reasoning skills and safety to developers. They help research, innovation, and transparent AI development, especially for smaller groups or emerging markets. By making powerful tools accessible, they support global progress and democratic use of AI. This encourages collaboration and new ideas for a better future.
Amid this release, rumors are intensifying about OpenAI's potential launch of GPT-5 today, with reports of a livestream at 10AM PT teasing major advancements, sparking widespread excitement and speculation about even more powerful AI capabilities that could redefine the field.