OpenAI has unveiled GPT-OSS, a fully open-weight AI model available for free download, customization, and local use even on a laptop.
GPT-OSS comes in two configurations: a 120-billion-parameter model that can run on a single Nvidia GPU and delivers performance comparable to o4-mini, and a 20-billion-parameter model that rivals o3-mini and operates on just 16GB of VRAM. Both versions are available today via Hugging Face, Databricks, Azure, and AWS under the Apache 2.0 license, enabling unrestricted modification and commercial use.
The move follows CEO Sam Altman’s January admission that OpenAI had been “on the wrong side of history” by avoiding open releases, especially as open-source AI models like DeepSeek gained traction for their cost efficiency and customization potential.
OpenAI says GPT-OSS can handle reasoning tasks, web browsing, coding, and agent operations via the company’s APIs. The model is text-only and keeps its training data undisclosed but exposes its chain of thought to help researchers monitor for misuse, deception, or unsafe behaviors. According to the company, GPT-OSS underwent the most extensive safety testing of any OpenAI release, including evaluations by external security firms to ensure minimal risks in sensitive areas like cybersecurity and biosecurity.
While OpenAI has not published benchmark comparisons against rivals such as Meta’s Llama, Google’s Gemma, or DeepSeek, it claims both GPT-OSS variants match its closed reasoning models on coding and reasoning tests, including Humanity’s Last Exam. “These are incredible models,” said OpenAI co-founder Greg Brockman. “The team really cooked with this one.”
OpenAI isn’t committing to a fixed update schedule for GPT-OSS, but aims for it to empower smaller developers and startups seeking greater control over their AI tools. “Lower the barrier to access, and innovation goes up,” Brockman said. “Let people hack, and they’ll surprise you.”
