Despite what its name suggests, OpenAI hadn’t released an “open” model—one that includes access to the weights, or the numerical parameters often described as the model’s brains—since GPT-2 in 2020. That changed on Tuesday: The company launched a long-awaited open-weight model, in two sizes, that OpenAI says pushes the frontier of reasoning in open-source AI.
“We’re excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible,” said OpenAI CEO Sam Altman about the release. “As part of this, we are quite hopeful that this release will enable new kinds of research and the creation of new kinds of products.” He emphasized that he is “excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
Altman had teased the upcoming models back in March, two months after admitting, in the wake of the success of China’s open models from DeepSeek, that the company had been “on the wrong side of history” when it came to opening up its models to developers and builders. But while the weights are now public, experts note that OpenAI’s new models are hardly “open.” By no means is it giving away its crown jewels: The proprietary architecture, routing mechanisms, training data, and methods that power its most advanced models—including the long-awaited GPT-5, widely expected to be released sometime this month—remain tightly under wraps.
OpenAI is targeting AI builders and developers
The two new model names—gpt-oss-120b and gpt-oss-20b—may be indecipherable to non-engineers, but that’s because OpenAI is setting its sights on AI builders and developers seeking to rapidly build on real-world use cases on their own systems. The company noted that the larger of the two models can run on a single Nvidia 80GB chip, while the smaller one fits on consumer hardware like a Mac laptop.
Greg Brockman, cofounder and president of OpenAI, acknowledged on a press pre-briefing call that “it’s been a long time” since the company had released an open model. He added that it is “something that we view as complementary to the other products that we release” and along with OpenAI’s proprietary models, “combine to really accelerate our mission of ensuring that AGI benefits all of humanity.”
OpenAI said the new models perform well on reasoning benchmarks, which have emerged as the key measurements for AI performance, with models from OpenAI, Anthropic, Google and DeepSeek fiercely competing over their abilities to tackle multistep logic, code generation, and complex problem-solving. Ever since the open source DeepSeek R1 shook the industry in January with its reasoning capabilities at a much lower cost, many other Chinese models have followed suit—including Alibaba’s Qwen and Moonshot AI’s Kimi models. While OpenAI said at a press pre-briefing that the new open-weight models are a proactive effort to provide what users want, it is also clearly a strategic response to ramping up open-source competition.
Notably, OpenAI declined to benchmark its new open-weight models against Chinese open-source systems like DeepSeek or Qwen—despite the fact that those models have recently outperformed U.S. rivals on key reasoning benchmarks. In the press briefing, the company said it is confident in its benchmarks against its own models and that it would leave it to others in the AI community to test further and “make up their own minds.”
Avoiding the leak of intellectual property
OpenAI’s new open-weight models are built using a mixture-of-experts (MoE) architecture, in which the system activates only the “experts,” or subnetworks, it needs for a specific input, rather than using the entire model for every query. Dylan Patel, founder of research firm SemiAnalysis, pointed out in a post on X before the release that OpenAI trained the models only using publicly known components of the architecture—meaning the building blocks it used are already familiar to the open-source community. He emphasized that this was a deliberate choice—that by avoiding any proprietary training techniques or architecture innovations, OpenAI could release a genuinely useful model without actually leaking any intellectual property that powers its proprietary frontier models like GPT-4o.
For example, in a model card accompanying the release, OpenAI confirmed that the models use a mixture-of-experts (MoE) architecture with 12 active experts out of 64, but it does not describe the routing mechanism, which is a crucial and proprietary part of the architecture.
“You want to minimize risk to your business, but you [also] want to be maximally useful to the public,” Aleksa Gordic, a former Google DeepMind researcher, told Fortune, adding that companies like Meta and Mistral, which have also focused on open-weight models, have similarly not included proprietary information.
“They minimize the IP leak and remove any risk to their core business, while at the same time sharing a useful artifact that will enable the startup ecosystem and developers,” he said. “It’s by definition the best they can do given those two opposing objectives.”
