OpenAI has released two new open-weight language models, marking a major shift in its strategy. The models, named gpt-oss-120b and gpt-oss-20b, can be run locally on personal devices and fine-tuned for different needs. This is the first open-weight release from OpenAI since GPT-2 came out in 2019.
CEO Sam Altman said the move is aimed at putting powerful AI tools into more hands. "We're excited to make this model, the result of billions of dollars of research, available to the world," he said in a statement posted on X.
The models are now available for download on Hugging Face, a platform that hosts AI tools and libraries.
What is an open-weight model?
Unlike traditional models such as ChatGPT, which are accessed through online services, open-weight models allow users to see the internal workings of the AI. This includes the “weights” or parameters that shape how the model processes information.
OpenAI co-founder Greg Brockman described the release as “complementary” to their paid services. He stated that these models offer different strengths, such as offline use and local customisation. For instance, gpt-oss-20b can run on consumer-grade devices with at least 16 GB of memory.
These models use a chain-of-thought reasoning approach, meaning they go through step-by-step reasoning before giving answers. While they do not support images or voice inputs, they can browse the web, execute code, and work as AI agents in software environments.
What can developers do with them?
Both models are being released under the Apache 2.0 licence, which allows for commercial use and redistribution. This brings OpenAI in line with other major releases like Mistral and Qwen by Alibaba.
Trending Stories
Eric Wallace, a safety researcher at OpenAI, said the models were tested against potential misuse scenarios. “We actually fine-tuned the model internally on some of these risk areas,” he told reporters. Based on OpenAI’s risk framework, the models did not show high-risk behaviour.
Chris Koch, another researcher at OpenAI, said gpt-oss-120b performs at a level similar to the company's proprietary models, including o3 and o4-mini, and sometimes even better in benchmark tests.

&imwidth=800&imheight=600&format=webp&quality=medium)
)
)
)
)
)
)
)
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
&im=FitAndFill=(700,400))
&im=FitAndFill=(700,400))
)
)
)
&im=FitAndFill=(700,400))
)
)
)
&im=FitAndFill=(700,400))
&im=FitAndFill=(700,400))
&im=FitAndFill=(700,400))
)
)
)
)