For the first time since GPT-2 in 2019, OpenAI is releasing new open-weight large language models. It’s a major milestone for a company that has increasingly been accused of forgoing its original stated mission of “ensuring artificial general intelligence benefits all of humanity.” Now, following multiple delays for additional safety testing and refinement, gpt-oss-120b and gpt-oss-20b are available to download from Hugging Face.
Before going any further, it’s worth taking a moment to clarify what exactly OpenAI is doing here. The company is not releasing new open-source models that include the underlying code and data the company used to train them. Instead, it’s sharing the weights — that is, the numerical values the models learned to assign to inputs during their training — that inform the new systems. According to Benjamin C. Lee, professor of engineering and computer science at the University of Pennsylvania, open-weight and open-source models serve two very different purposes.
“An open-weight model provides the values that were learned during the training of a large language model, and those essentially allow you to use the model and build on top of it. You could use the model out of the box, or you could redefine or fine-tune it for a particular application, adjusting the weights as you like,” he said. If commercial models are an absolute black box and an open-source system allows for complete customization and modification, open-weight AIs are somewhere in the middle.
OpenAI has not released open-source models, likely since a rival could use the training data and code to reverse engineer its tech. “An open-source model is more than just the weights. It would also potentially include the code used to run the training process,” Lee said. And practically speaking, the average person wouldn’t get much use out of an open-source model unless they had a farm of high-end NVIDIA GPUs running up their electricity bill. (They would be useful for researchers looking to learn more about the data the company used to train its models though, and there are a handful of open-source models out there like Mistral NeMo and Mistral Small 3.)
With that out of the way, the primary difference between gpt-oss-120b and gpt-oss-20b is how many parameters each one offers. If you’re not familiar with the term, parameters are the settings a large language model can tweak to provide you with an answer. The naming is slightly confusing here, but gpt-oss-120b is a 117 billion parameter model, while its smaller sibling is a 21-billion one.
In practice, that means gpt-oss-120b requires more powerful hardware to run, with OpenAI recommending a single 80GB GPU for efficient use. The good news is the company says any modern computer with 16GB of RAM can run gpt-oss-20b. As a result, you could use the smaller model to do something like vibe code on your own computer without a connection to the internet. What’s more, OpenAI is making the models available through the Apache 2.0 license, giving people a great deal of flexibility to modify the systems to their needs.
Despite this not being a new commercial release, OpenAI says the new models are in many ways comparable to its proprietary systems. The one limitation of the oss models is that they don’t offer multi-modal input, meaning they can’t process images, video and voice. For those capabilities, you’ll still need to turn to the cloud and OpenAI’s commercial models, something both new open-weight systems can be configured to do. Beyond that, however, they offer many of the same capabilities, including chain-of-thought reasoning and tool use. That means the models can tackle more complex problems by breaking them into smaller steps, and if they need additional assistance, they know how to use the web and coding languages like Python.
Additionally, OpenAI trained the models using techniques the company previously employed in the development of o3 and its other recent frontier systems. In competition-level coding gpt-oss-120b earned a score that is only a shade worse than o3, OpenAI’s current state-of-the-art reasoning model, while gpt-oss-20b landed in between o3-mini and o4-mini. Of course, we’ll have to wait for more real-world testing to see how the two new models compare to OpenAI’s commercial offerings and those of its rivals.
The release of gpt-oss-120b and gpt-oss-20b and OpenAI’s apparent willingness to double down on open-weight models comes after Mark Zuckerberg signaled Meta would release fewer such systems to the public. Open-sourcing was previously central to Zuckerberg’s messaging about his company’s AI efforts, with the CEO once remarking about closed-source systems “fuck that.” At least among the sect of tech enthusiasts willing to tinker with LLMs, the timing, accidental or not, is somewhat embarrassing for Meta.
“One could argue that open-weight models democratize access to the largest, most capable models to people who don’t have these massive, hyperscale data centers with lots of GPUs,” said Professor Lee. “It allows people to use the outputs or products of a months-long training process on a massive data center without having to invest in that infrastructure on their own. From the perspective of someone who just wants a really capable model to begin with, and then wants to build for some application. I think open-weight models can be really useful.”
OpenAI is already working with a few different organizations to deploy their own versions of these models, including AI Sweden, the country’s national center for applied AI. In a press briefing OpenAI held before today’s announcement, the team that worked on gpt-oss-120b and gpt-oss-20b said they view the two models as an experiment; the more people use them, the more likely OpenAI is to release additional open-weight models in the future.
