OpenAI releases open-weight “gpt-oss” models to compete with DeepSeek, Meta

Today, OpenAI released “gpt-oss,” its first “open-weight” large language model since 2019’s GPT-2. The new “reasoning” model comes in two sizes, 120b and 20b, and can be run locally on a laptop — and the smaller one can even run on a phone, according to the company. OpenAI says the new models outperform or exceed the company’s proprietary o3, o3-mini, and 04-mini in some tasks. The models do not generate images or video.

The models are open-source, which means free for anyone to use — but importantly, they are also “open-weight,” which means the internal parameters generated from training the model are made available. Open-weight models, such as Meta’s Llama models, allow developers to customize the model further or run them on their own infrastructure without having to pay OpenAI a license or subscription.

The release of DeepSeek’s open-source model, which shook the AI world with its faster, cheaper, better approach to doing more with less, put pressure on companies like OpenAI that mainly offered proprietary models that had to run on their own infrastructure.

Releasing the model weights does not include the original training data used by the developers. OpenAI does not share the models or weights for its recent models, including GPT-3, GPT-4, or its “o” series models.

In a post on X announcing the new models, OpenAI cofounder and CEO Sam Altman said they are a “big deal” and that the company is hopeful the release will “enable new kinds of research and the creation of new kinds of products. We expect a meaningful uptick in the rate of innovation in our field, and for many more people to do important work than were able to before.”

“We believe far more good than bad will come from it.”

OpenAI says that it dedicates a huge amount of resources to making sure its hosted models are safe from misuse, but releasing a powerful open-weight model that’s on par with its current state-of-the-art models creates new risks beyond the visibility of OpenAI’s safety team.

Altman wrote, “We have worked hard to mitigate the most serious safety issues, especially around biosecurity. gpt-oss models perform comparably to our frontier models on internal safety benchmarks.”

Acknowledging the balancing of risks and benefits of releasing such a powerful technology, Altman wrote, “We believe far more good than bad will come from it.”