OpenAI has finally released its first open-weight large language models since 2019’s GPT-2. These new “gpt-oss” models are available in two different sizes and score similarly to the company’s o3-mini and o4-mini models on several benchmarks. Unlike the models available through OpenAI’s web interface, these new open models can be freely downloaded, run, and even modified on laptops and other local devices.
In the company’s many years without an open LLM release, some users have taken to referring to it with the pejorative “ClosedAI.” That sense of frustration had escalated in the past few months as these long-awaited models were delayed twice—first in June and then in July. With their release, however, OpenAI is reestablishing itself as a presence for users of open models.
That’s particularly notable at a time when Meta, which had previously dominated the American open-model landscape with its Llama models, may be reorienting toward closed releases—and when Chinese open models, such as DeepSeek’s offerings, Kimi K2, and Alibaba’s Qwen series, are becoming more popular than their American competitors.
“The vast majority of our [enterprise and startup] customers are already using a lot of open models,” said Casey Dvorak, a research program manager at OpenAI, in a media briefing about the model release. “Because there is no [competitive] open model from OpenAI, we wanted to plug that gap and actually allow them to use our technology across the board.”
The new models come in two different sizes, the smaller of which can theoretically run on 16 GB of RAM—the minimum amount that Apple currently offers on its computers. The larger model requires a high-end laptop or specialized hardware.
Open models have a few key use cases. Some organizations may want to customize models for their own purposes or save money by running models on their own equipment, though that equipment comes at a substantial upfront cost. Others—such hospitals, law firms, and governments—might need models that they can run locally for data security reasons.
OpenAI has facilitated such activity by releasing its open models under a permissive Apache 2.0 license, which allows the models to be used for commercial purposes. Nathan Lambert, post-training lead at the Allen Institute for AI, says that this choice is commendable: Such licenses are typical for Chinese open-model releases, but Meta released its Llama models under a bespoke, more restrictive license. “It’s a very good thing for the open community,” he says.
Researchers who study how LLMs work also need open models, so that they can examine and manipulate those models in detail. “In part, this is about reasserting OpenAI’s dominance in the research ecosystem,” says Peter Henderson, an assistant professor at Princeton University who has worked extensively with open models. If researchers do adopt gpt-oss as new workhorses, OpenAI could see some concrete benefits, Henderson says—it might adopt innovations discovered by other researchers into its own model ecosystem.
More broadly, Lambert says, releasing an open model now could help OpenAI reestablish its status in an increasingly crowded AI environment. “It kind of goes back to years ago, where they were seen as the AI company,” he says. Users who want to use open models will now have the option to meet all their needs with OpenAI products, rather than turning to Meta’s Llama or Alibaba’s Qwen when they need to run something locally.
The rise of Chinese open models like Qwen over the past year may have been a particularly salient factor in OpenAI’s calculus. An employee from OpenAI emphasized at the media briefing that the company doesn’t see these open models as a response to actions taken by any other AI company, but OpenAI is clearly attuned to the geopolitical implications of China’s open-model dominance. “Broad access to these capable open-weights models created in the US helps expand democratic AI rails,” the company wrote in a blog post announcing the models’ release.
Since DeepSeek exploded onto the AI scene at the start of 2025, observers have noted that Chinese models often refuse to speak about topics that the Chinese Communist Party has deemed verboten, such as Tiananmen Square. Such observations—as well as longer-term risks, like the possibility that agentic models could purposefully write vulnerable code—have made some AI experts concerned about the growing adoption of Chinese models. “Open models are a form of soft power,” Henderson says.
Lambert released a report on Monday documenting how Chinese models are overtaking American offerings like Llama and advocating for a renewed commitment to domestic open models. Several prominent AI researchers and entrepreneurs, such as HuggingFace CEO Clement Delangue, Stanford’s Percy Liang, and former OpenAI researcher Miles Brundage, have signed on.
The Trump administration, too, has emphasized development of open models in its AI Action Plan. With both this model release and previous statements, OpenAI is aligning itself with that stance. “In their filings about the action plan, [OpenAI] pretty clearly indicated that they see US–China as a key issue and want to position themselves as very important to the US system,” says Rishi Bommasani, a senior research scholar at the Stanford Institute for Human-Centered Artificial Intelligence.
And OpenAI may see concrete political advantages from aligning with the administration’s AI priorities, Lambert says. As the company continues to build out its extensive computational infrastructure, it will need political support and approvals, and sympathetic leadership could go a long way.
