OpenAI is releasing a new open-weight model dubbed GPT-OSS that can be downloaded for free, be customized, and even run on a laptop.
The model comes in two variants: 120-billion-parameter and 20-billion-parameter versions. The bigger version can run on a single Nvidia GPU and performs similarly to OpenAI’s existing o4-mini model, while the smaller version performs similarly to o3-mini and runs on just 16GB of memory. Both model versions are being released today via platforms like Hugging Face, Databricks, Azure, and AWS under the Apache 2.0 license, which allows them to be widely modified for commercial purposes.
This is OpenAI’s first open-weight model in over six years, years before the debut of ChatGPT. Until earlier this year, CEO Sam Altman cited safety concerns as the main reason for not releasing a follow-up. Meanwhile, developers have flocked to open models due to their lower cost and customizability. In January, after the rise of DeepSeek, Altman said that OpenAI had “been on the wrong side of history” by not releasing its own open models.
Now, OpenAI is reasserting itself with an open-weight model that it says can perform reasoning tasks, browse the web, write code, and operate agents via the company’s existing APIs. “I think a lot of people are actually surprised to know that the vast majority of our customers are already using a lot of open models,” Chris Cook, an OpenAI researcher, said during a media briefing. “We wanted to plug that gap and allow them to use our technology across the board.”
On the safety front, OpenAI says that GPT-OSS is its most rigorously tested model to date, and that it was tested with external safety firms to ensure it doesn’t pose risks in areas like cybersecurity and biological weapons. The model’s chain of thought, or visible process used to arrive at an answer, is shown “to monitor model misbehavior, deception and misuse,” according to a company press release. Its output is text-only and, like all of OpenAI’s models, GPT-OSS’s training data is undisclosed.
“The team really cooked with this one.”
OpenAI hasn’t shared benchmarks comparing GPT-OSS to other open models like Llama, DeepSeek, or Google’s Gemma. Both variants of GPT-OSS perform similarly to OpenAI’s closed reasoning models on coding tasks and tests like Humanity’s Last Exam. “These are incredible models,” said OpenAI cofounder Greg Brockman. “The team really cooked with this one.”
OpenAI isn’t committing to a release schedule for future versions of GPT-OSS, but it hopes that the model will be used by smaller developers and companies that want more control over how their data is used. “We’ve always believed that if you lower the barrier to access, then innovation just goes up,” said Brockman. “You let people hack, then they will do things that are incredibly surprising.”
