Sam Altman revealed today that Optai will publish an open -weight artificial intelligence model in the coming months.
“We are delighted to publish a new powerful open language model with reasoning in the coming months,” wrote the CEO on X.
This decision is partly a response to the flight success of the R1 model of the Chinese company Deepseek, as well as the popularity of the Meta Llama models.
Shortly after the release of the Deepseek model in January, Altman said that Optai was “on the wrong side of history” concerning open models, reporting a likely change of direction. On Monday, he said that the company was planning to publish an open weight model for some time, adding “now it is important to do so”.
Openai may feel the need to show that it can form the new model at a lower cost, because the Deepseek model has been allegedly trained in a fraction of the cost of most large AI models.
“This is incredible news,” said Wired Wired, Clement Dengue, Co -founder and CEO of Huggingface, a company specializing in the accommodation of open AI models. “With Deepseek, everyone achieves the power of open weights.”
Openai is currently putting its AI available via a chatbot and via the cloud. R1, Llama and other open models can be downloaded for free and modified. The weights of a model refer to values within a large neural network – something that is defined during training. Open weight models are cheaper to use and can also be adapted to sensitive use cases, such as managing highly confidential information.
Steven Heidel, a member of OPENAI technical staff, republished Altman’s announcement and added: “We are publishing a model this year that you can execute on your own equipment.”
Johannes Heidecke, a researcher working on AI security in Openai, also republished the message on X, adding that the company would carry out rigorous tests to ensure that the open weight model could not easily be used in bad guys. Some AI researchers fear that open models can help criminals launching cyber attacks or even developing organic or chemical weapons. “Although open models bring unique challenges, we are guided by our preparation framework and will not publish models, we think we are pose catastrophic risks,” wrote Heidecke.
OPENAI TODAY has also published a web page inviting developers to request early access to the next model. Altman said in his article that the company would host events for developers with early prototypes of the new model in the coming weeks.
Meta was the first big company in AI to continue a more open approach, publishing the first version of Llama in July 2023. An increasing number of open -weight AI models are now available. Some researchers note that the Lama and some other models are not as transparent as they could be, because training data and other details are always secret. Meta also imposes a license that limits the ability of other companies to take advantage of the applications and tools built using Lama.
Update of March 31, 2025, 4:21 is: This article was updated with a comment from Clement Delangue, Co -founder and CEO of Huggingface.
