OpenAI is about to launch an open weight model: or next week, similar to O3 Mini
According to “theverge ” , informed sources, OpenAi is planning to launch an “open-weight model” (open-weight model), which will most likely be available next week, 17 July 2025. The open weight is an intermediate form between open and closed sources, and given that previously OpenAI’s models usually used the “closed source weight” model, i.e. training parameters were not open to the public, this initiative marked the first attempt by OpenAI to publish such a publication since 2019 when it was GPT-2.
According to the source, the performance of the new model is “similar to o3 Mini” and has a well-rated reasoning capability for the latest OpenAI model. o3 Mini is a highly effective reasoning model for the optimization of STEM (science, technology, engineering, mathematics) that is known for cost-effectiveness and rapid response. Unlike previous models, the new models will allow companies and governments to operate on their own, similar to the rapid adoption of DeepSeek R1 by cloud providers such as Microsoft earlier this year. The model is expected to be available on several cloud platforms such as Azure, Hugging Face. In recent months, OpenAI has shown the model to developers and researchers, and has publicly consulted the AI community to demonstrate its desire to optimize the model through community collaboration. Although sources claim that it may be on line next week, OpenAI ‘ s release dates are often adjusted by development challenges, server capacity or competitive dynamics, and the exact timing needs to be officially confirmed.
OpenAI ‘ s partnership with Microsoft is currently being renegotiated, as OpenAI is seeking to restructure its corporate form into a profit-oriented entity. The sole-source agreement signed by the two companies in 2023, which grants Microsoft the right to resell and promote the OpenAI model through the Azure OpenAI service, has become the backbone of the corporate AI strategy of both companies. However, the emergence of viable open-source and open-weight models, including Llama 2 in Meta and DeepSeek R1, has changed the competitive trend and forced OpenAI to reconsider which technologies must remain closed. In addition, Microsoft itself is constantly refining its cooperation agreement with OpenAI. Although initially exclusive, Microsoft agreed this year to allow OpenAI to use its competitor’s AC resources, such as Oracle, an initiative that was limited to modelling at the time, but now seems to signal greater openness.
Perhaps the most difficult question surrounding the forthcoming version of OpenAI is how it will be “open” in practice. The term “open model” is often confused with “open source” — two concepts that overlap and are distinct. The “open weight” model will support local and third-party hosting by issuing trained neural network parameters. But the real open source also requires full access to model codes, training programmes, data management details. Ideally, there is also a need for a liberal permit that allows for modification, redistribution and even commercial use. OpenAI’s past behaviour also increased uncertainty. Although GPT-2 was finally fully open in 2019, its state-of-the-art model remains proprietary. The open weight model, which is likely to be presented for the first time this month, is expected to follow a stricter path, similar to the Llama 2 licence in Meta, which prohibits certain usages and limits redistribution in specific circumstances.
