After months of requests by developers, and plenty of trolling from the larger AI community, OpenAI has finally released an open model.
OpenAI has released two open-weights models. The GPT-oss-120b is a large open model designed to run in data centers and on high-end laptops and desktops. The GPT-oss-20b is a medium-sized open model that can run on most desktops and laptops. The models are supported by the Apache 2.0 license, allowing developers to build freely on them. Both models are reasoning models.
“Gpt-oss is a big deal; it is a state-of-the-art open-weights reasoning model, with strong real-world performance comparable to o4-mini, that you can run locally on your own computer (or phone with the smaller size). We believe this is the best and most usable open model in the world,” OpenAI CEO Sam Altman said on the release.
The models are designed for agentic tasks, and leverage powerful instruction following, including web search and Python code execution. Users can adjust the reasoning effort in the models to low, medium or high. The models can also be customized to specific use-cases with full-parameter finetuning.
In a demo, OpenAI showed off the model being run on a laptop, and controlling the user’s screen.
The models, predictably, underperform o3 and o4-mini across most benchmarks. The larger gpt-oss-120b model, though, beats o4-mini on the Humanity’s Last Exam benchmark.
OpenAI says it has run safety tests on the models to ensure that they aren’t misused. The potential for misuse in open-source models is much greater than OpenAI’s closed-source models, and OpenAI says it has taken steps to prevent this misuse.
The release of the open-weights models will help OpenAI blunt a particular criticism against itself — while the company had been named “OpenAI”, originally as an open alternative to Google, it hadn’t released any of its model weights since GPT-2. Meanwhile, a host of open-weights models had cropped up, including Meta’s Llama, China’s DeepSeek, and more recently Kimi K2. These models had seen plenty of adoption from developers, because they could fine-tune the models to their own needs. OpenAI, too, appears to have finally joined the open race with an open-weights model of its own. It remains to be seen how the models are perceived by users and the developer community, but given the benchmark scores and the buzz around OpenAI in general, one would expect they end up becoming quite popular.