OpenAI has released its first open-weight models in five years: gpt-oss-120b and gpt-oss-20b. These powerful new models are designed for advanced reasoning and can run on personal devices, making advanced AI more accessible for everyone.
OpenAI has published two new strong language models, its first public open-weight release in more than five years. The models, gpt-oss-120b and gpt-oss-20b, are built to be experts in complex reasoning and are available for free download, marking an intentional change towards an open and open-access ecosystem for AI creation.
This launch is a break from OpenAI's recent emphasis on closed-source, proprietary models and a return to its original values of putting AI in the service of all of humanity.
OpenAI CEO, Sam Altman, said
“We're excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the many people as possible”.
The models are now hosted on Hugging Face and available under the permissive Apache 2.0 license, allowing for broad use, including for commercial purposes.
Key Features and Capabilities
The new models are distinguished by their advanced reasoning abilities, which are crucial for complex problem-solving. They are particularly adept at tasks involving coding, competitive mathematics, and health-related inquiries. They are well suited to code-related tasks, competitive math, and health-related questions.
Perhaps the standout feature is the models' capacity for "chain-of-thought" reasoning, a method that solves problems in several steps, giving users better insight into how the AI came to its answer.
Importantly, these models are optimized for efficiency. The larger model, gpt-oss-120b, can be run on a single graphics processing unit (GPU), while its smaller counterpart, gpt-oss-20b, is compact enough to run on a standard consumer laptop with sufficient memory.
This local operation capability is a key differentiator, as it allows users to run the AI offline and behind their firewalls, addressing critical concerns around data privacy and security.
A Strategic Move in a Contested Landscape
OpenAI's re-entry into the open-weight space comes amid a highly competitive and politically charged environment. For some time, other companies, notably Meta with its Llama models and China's DeepSeek, have been leaders in this sector. OpenAI's release is seen as a direct effort to regain leadership and promote what it calls "US-led democratic AI."
By making its models widely available, OpenAI aims to establish a global standard for AI development rooted in reliability, transparency, and democratic norms. The company's “OpenAI for Countries” initiative and support for nonprofits are integral to this vision, providing advanced AI infrastructure to allies and grassroots organizations alike. This approach is viewed as a form of "soft power," where the most widely adopted AI models can influence global standards.
OpenAI Open-Weight Models Launch on AWS for Global Use
The release has also spurred a new partnership with Amazon, which is now offering the GPT-oss models on its Bedrock platform. This first-time collaboration provides millions of AWS customers with access to powerful, open-source AI. The partnership is a win for both sides: it expands the reach of OpenAI's advanced technology to AWS's vast customer base and reinforces AWS's commitment to offering a broad choice of models, including those that are highly price-performant compared to competitors.
Atul Deo, director of product, AWS, said on this partnership,
"The addition of OpenAI as our newest open weight model provider marks a natural progression in our commitment to bringing cutting-edge AI to organizations worldwide, and the unmatched size of our customer base marks a transformative shift in access to OpenAI’s advanced technology."
Safety and the Future of AI
OpenAI recognized the inherent safety risks in releasing open-weight models, which can be fine-tuned for malicious use. In response, the company performed extensive testing, including simulating how the models could be used for harm by malicious users.
The models, said a safety researcher, were fine-tuned specifically to mitigate risks in areas of concern. The company's evaluation, according to its "preparedness framework," determined that the models do not present a high degree of risk.
The release by the company of gpt-oss-120b and gpt-oss-20b is a turning point, not only for OpenAI, but for the entire AI sector.
With its aim to strike a balance between proprietary innovation and an openness policy of making the technology universally accessible, OpenAI is striving to define the next generation of AI development, spreading the benefits of the technology far and wide while keeping itself and America ahead in the run.
