OpenAI has introduced two new AI models, gpt-oss-120b and gpt-oss-20b, intended to help developers run complex systems more affordably. These are open-weight models, meaning the internal data used to train them is available to the public. Anyone with the interest and the means, be they engineers or academics, can download the models, adjust them, and use them as they see fit.
The announcement was made on X by OpenAI’s chief executive, Sam Altman, who wrote that “gpt-oss is out! We made an open model that performs at the level of o4-mini and runs on a high-end laptop (WTF!!) (and a smaller one that runs on a phone). super proud of the team; big triumph of technology.”
Empowering Developers and Researchers With Open Access AI
In a separate post, Altman wrote that the company supports individual freedom. While most users, he said, are likely to prefer the ease of tools like ChatGPT, people ought to have the option to shape and run their own AI systems when needed. The benefits for privacy, he noted, are self-evident.
He went on to say that the release may lead to fresh lines of research and entirely new types of products. In his view, it is likely to quicken the pace of innovation and widen the field, making it possible for more people to carry out valuable work than before.
The gpt-oss-120b and gpt-oss-20b models are made available under the Apache 2.0 license. The former, a larger model, is said to perform on par with OpenAI’s o4-mini, which remains a proprietary system. The smaller variant is intended for devices with more modest hardware, including mobile phones. Both are tailored for reasoning tasks and suited to applications such as conversational agents, code generation, and other intelligent tools.
Technical Specs and Key Features of GPT-OSS-120B and GPT-OSS-20B
According to OpenAI, the gpt-oss-120b model is able to run on a single GPU with 80GB of memory, while the smaller gpt-oss-20b has been built to function on systems with as little as 16GB. Both models are said to handle reasoning tasks, call functions, and operate tools. They are also designed to adjust how much effort they spend on reasoning, depending on the task, so as to strike a balance between speed and output.
The company stated that it has applied its internal safety methods to train and test these systems. It also carried out checks using a specially fine-tuned version of the larger model, shaped to withstand adversarial testing, to ensure it met the requirements set by its Preparedness Framework for safe AI deployment.
The training process relied on large collections of text, with a focus on computer code, scientific and technical material, and broad general knowledge. OpenAI partnered with early testers, among them AI Sweden and Orange, to explore uses such as secure, on-site deployment.
Both models are built using the Transformer design and make use of modern methods to reduce memory use. The 120b version contains 117 billion parameters, while the 20b model holds 21 billion.
Final Words
The launch of gpt-oss-120b and gpt-oss-20b has effectively given the keys to the AI kingdom to anyone with a halfway decent laptop and a vision. Whereas the rest of us are likely to remain content with inquiring ChatGPT on the dinner recipes and resolving online disputes, the open-source models provide something more substantial: the chance to fiddle with the engine without the corporate watchdog snapping at your heels.
The idea of democratized AI development that Altman is proposing sounds idealistic, but it is hard not to think that we are seeing the emergence of a thousand garage-built AI experiments or just a new way of people automating their homework. In any case, the shift represents a radical change of direction in the approach of OpenAI, which was previously focused on keeping their cards close to their chests to now giving everyone a hand. It is yet to be seen whether this gamble will pay off but the ante has definitely been upped in the AI poker game.
