Toward the end of March, OpenAI announced its plan to launch a new “open” language model following GPT-2 later this year. Now, more information about the model is starting to emerge from the company’s interactions with the AI development community.

Aidan Clark, VP of research at OpenAI, is spearheading the development of this open model, which is currently in its early stages as per sources familiar with the project. The company aims to unveil the model, a reasoning model similar to its o-series models, in early summer and ensure it surpasses benchmarks set by other open reasoning models.

OpenAI is considering a highly permissive license for the model with minimal usage or commercial restrictions, sources reveal. Models like Llama and Google’s Gemma have faced criticism for their demanding requirements, which OpenAI seems to be trying to avoid.

The ChatGPT developer is under pressure from competitors like DeepSeek, a Chinese AI lab that takes an open approach to model launches. Unlike OpenAI, these “open” rivals share their models with the AI community for experimentation and, in some cases, commercial use.

Success stories have emerged from this strategy. Meta, focusing on its Llama AI models, reported over 1 billion downloads in early March. DeepSeek also quickly garnered a large user base worldwide and attracted investors.

Sources indicate that OpenAI plans for its open model, which will involve “text in, text out,” to be compatible with high-end consumer hardware. Developers may be able to switch its “reasoning” on or off, similar to models from other companies like Anthropic. If well-received, OpenAI may release additional models, possibly smaller ones as well.

OpenAI’s CEO Sam Altman has previously expressed the need for a different open-source strategy, acknowledging the importance of improving future models while ensuring a smaller lead in the industry.

Regarding safety measures, Altman stated that the upcoming open model will undergo rigorous evaluation and red-teaming. The company intends to release a model card, offering detailed technical insights and safety testing results.

OpenAI has faced criticism from AI ethicists for hastily conducting safety tests on recent models and failing to provide model cards for some. Altman has also been accused of misinforming OpenAI executives about model safety reviews before a brief departure in November 2023.

We have reached out to OpenAI for comments and will provide updates if we receive a response.