For the first time since 2019, OpenAI has released an open weight model.
On Tuesday, OpenAI launched gpt-oss-120b and gpt-oss-20b, two open weight language models available under the Apache 2.0 license.
According to OpenAI, the models demonstrate strong tool use capabilities and are optimized for deployment on consumer hardware. The AI vendor said the models were trained using a mix of reinforcement learning techniques informed by OpenAI's other models, such as o3.
Compared with core reasoning models, gpt-oss-120b is similar to OpenAI's o4-mini, while the gpt-oss-20b model is similar to o3-mini. The 20b model can also run on edge devices with 16 GB of memory, OpenAI said.
OpenAI trained the models on Nvidia H100 GPUs and optimized them to run inference across the Nvidia stack, including on Nvidia Blackwell and RTX GPUs, Nvidia said.
Anticipated release
OpenAI's open weight models were highly anticipated. They come after heavy criticism from the AI community that the pioneering generative AI vendor has veered far from its original path, when it used to produce open models.
Other GenAI providers have capitalized on the growing popularity of open source following the big splash of Chinese vendor DeepSeek's R1 model earlier this year.
Open weight models are models for which the training parameters are openly available. With open source models, coding sources and training data are available in addition to the parameters.
This is a response to a lot of the open source models out there.
Lian Jye SuAnalyst, Omdia
"This is a response to a lot of the open source models out there," said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. Many vendors in the U.S. and China have released open models, including Google, Alibaba, IBM, Ai2 and others, he noted.
"It's about time," Su said. "It's good to see them contributing to the open source community after being absent for a long time."
Su said the models themselves are typical compared with other models in the market in that OpenAI is using known techniques, such as a mixture of experts.
Some challenges and strategy
However, the models seem limited in terms of the languages they were trained on -- they were mostly trained in English. "They may have some problems working in other languages," said Rowan Curran, an analyst at Forrester Research.
The models' release is also significant after Meta CEO Mark Zuckerberg revealed on July 30 that the social media giant won't make its anticipated superintelligence AI models open. This could mean that Meta might no longer release new Meta Llama models. Llama is one of the most widely used open weight model families.
"It's a very nice moment to see this today," Curran said. He added that it's also useful that OpenAI released a larger version for data centers and a miniature version for the edge.
According to David Nicholson, an analyst at The Futurum Group, the release is also a strategy for OpenAI to target companies that are fine-tuning models. Because OpenAI released gpt-oss-120b and gpt-oss-20b under an Apache license, one could argue that the licensing is less restrictive and more open to commercial activity, he said.
"It's one thing to say that your model is open weight ... for academic purposes, but as soon as you're going to start making money, that's a whole different story," Nicholson said. "Apache is well known. Enterprises will be more comfortable with how these models are released and the licensing wrapper around them."
He added that the release also speaks to OpenAI's strategy in the tight GenAI race.
"This, from OpenAI, is a very strategic move," Nicholson continued. "It undermines Meta's attempt to creep into the commercial space."
He said it will be interesting to see the kind of market penetration OpenAI's open weight models achieve compared with Meta Llama models by the time Meta holds its next LLamaCon conference.
A call to do more
Despite the excitement about OpenAI's open weight models and how the vendor will contribute to the open source community, there is still a push for the ChatGPT creator to do more.
"The quality of their model seems good," said Hannaneh Hajishirzi, senior director of AI at Ai2 and a professor at the University of Washington.
She added that while the models will enable the open source community to build better models, OpenAI should release the models' data, full recipe, and process of what worked and didn't work at various stages of training. That would, in essence, make the models fully open source.
"I understand why they might not want to release all these details, because almost all the secrets are always in the details," Hajishirzi said. "This is an added benefit of what we mean by fully open."
As part of the release, Microsoft is bringing GPU-optimized versions of the gpt-oss-20b model to Windows. Developers who want to try the models can do so on OpenAI's open model playground and download them on Hugging Face. The models are also available on Together AI, and gpt-oss-120b is now live on Cerebras.
Esther Shittu is an Informa TechTarget news writer and podcast host covering AI software and systems.
