Toward the end of March, OpenAI announced plans to unveil its first “open” language model since the release of GPT-2, aiming for a release sometime this year. Details about this model are now emerging from the company’s discussions with the AI developer community.
According to sources from TechCrunch, the development of the open model is being led by Aidan Clark, OpenAI’s Vice President of Research, and is currently in its early stages. OpenAI aims to launch the model by early summer and intends for it to be a leading reasoning model among other open models, similar to OpenAI’s o-series.
The company is reportedly considering a highly permissive license for this model, with minimal usage or commercial restrictions. This approach contrasts with other open models like Llama and Google’s Gemma, which have faced criticism for their strict requirements. OpenAI appears to be aiming to avoid these criticisms.
Competition is mounting from other AI labs, such as the Chinese lab DeepSeek, which have taken an open approach to launching models, allowing availability for experimentation and commercialization within the AI community. This strategy has proven successful for some, with Meta’s Llama models having surpassed one billion downloads by early March, while DeepSeek has rapidly expanded its user base and attracted attention from investors.
Sources also indicate that OpenAI’s upcoming model will feature “text in, text out” functionality, compatible with high-end consumer hardware, and possibly allow developers to toggle its “reasoning” capabilities. This is similar to models released recently by Anthropic and others, which improve accuracy at the expense of latency. Should the launch be successful, OpenAI may consider releasing additional models, potentially including smaller ones.
In past statements, OpenAI CEO Sam Altman acknowledged the company may have been on the “wrong side of history” regarding open-sourcing its technology. He has expressed a desire to develop a new open-source strategy, although it is not currently a top priority for OpenAI. Altman mentioned that while they plan to create superior models, they may not maintain as significant a lead as in previous years.
Additionally, Altman has assured that the forthcoming open model will undergo thorough safety evaluations and red-teaming. TechCrunch’s sources report that OpenAI plans to release a model card for the open model, detailing its internal and external benchmarking and safety testing outcomes.
Altman stated in a post on X that the model will be evaluated using OpenAI’s preparedness framework, with additional scrutiny due to anticipated post-release modifications. However, the company has faced criticism from AI ethicists for reportedly expediting safety testing of past models and failing to release model cards for some. Altman has also been accused of misinforming OpenAI executives about safety reviews prior to his temporary removal as CEO in November 2023.
OpenAI has been contacted for comment, and any updates will be provided should they respond.
