OpenAI chief executive Sam Altman said the company is to release its first open-weight model since GPT-2 in the coming months, as it faces growing pressure from rivals that can be more easily customised by their users.

Open-weight models provide the ability to customise the weights, or relationships among the billions of parameters that are set during the model’s training, providing more visibility and control than closed systems.

Such models can generally be used on a company’s own hardware, meaning the model can handle a company’s sensitive data without it ever leaving the premises, a significant factor in areas such as banking or healthcare.

Customisation

The result is a customised AI model that is more under the organisation’s control.

Such offerings don’t go as far as open-source models, which give users access to a model’s source code and allow them to customise it.

Since GPT-2, released in 2019, OpenAI has been offering closed models, but the approach has come under pressure since the success in January of Chinese start-up DeepSeek, which releases its models and other data as open source.

Facebook parent Meta and Chinese tech giant Alibaba have had broad success with their Llama and Qwen open-source AI models.

Altman said the upcoming open-weight model will also have “reasoning” capabilities, a technique designed to reduce errors in output.

“We are excited to release a powerful new open-weight language model with reasoning in the coming months,” Altman said in a social media post.

He said the company had been considering releasing an open-weight model for some time and “now it feels important to do”.

Steven Heidel, a member of OpenAI’s technical staff, said in a separate post that the company would be “releasing a model this year that you can run on your own hardware”.

As open-weight models can be customised and used relatively independently, OpenAI is conducting testing to ensure the upcoming model can’t be used for harmful purposes, said safety researcher Johannes Heidecke.

‘Unique challenges’

“While open models bring unique challenges, we’re guided by our Preparedness Framework and will not release models we believe pose catastrophic risks,” he said.

OpenAI began accepting applications for developers to gain early access to the open-weight model and said it would be hosting developer events with early prototypes in the coming weeks in San Francisco, followed by Europe and Asia-Pacific.

“We still have some decisions to make, so we are hosting developer events to gather feedback and later play with early prototypes,” Altman said.

OpenAI is nearing a $40 billion (£31bn) funding round led by SoftBank, but to gain access to the full funding must complete its transition to a for-profit company by the end of 2025, according to multiple reports.

Excerpts of a forthcoming book about OpenAI by Wall Street Journal reporter Keach Hagey published over the weekend allege that Altman misled OpenAI’s board about model safety reviews ahead of his brief ouster in November 2023.

The allegations are similar to comments made by former board member Helen Toner in May of last year.