OpenAI will release a model with open scales – what does this mean?

Sam Altman announced on X (formerly Twitter) that OpenAI plans to release an open-weights model within a few months. This is different from the company’s previous approach, when models such as GPT-3.5 and GPT-4 were only available through the API. Until now, users could only work with these models through the OpenAI interface or by integrating API keys into their applications, and the models themselves remained closed and commercial.

However, models with open weights are a completely different approach. They allow you to publicly obtain the parameters of the trained model (the so-called «weights») that determine its answers. Developers and researchers can download these models, run them locally, adapt them to their own needs, or integrate them into custom applications. However, the openness of the weights does not mean that everything else will be published: the training code, datasets, and detailed logs may remain closed, so the model may not be fully open in the classic open-source sense.

Competitors such as Meta and Mistral have already released powerful open source models, including Llama and Mixtral. Other players, including DeepSeek, Alibaba, and Baichuan, also offer competitive solutions that in some cases even outperform closed models. Many developers and researchers prefer open scales because it provides more flexibility, transparency, and cost control.

To support the launch, OpenAI is planning a series of developer events in San Francisco, Europe, and Asia Pacific. During these sessions, users will be able to test prototypes, customize the model for their own tasks, and provide feedback.

Altman noted that OpenAI will evaluate the new model using its own security verification system, just like any other model. Since this product will be open and subject to modifications, the company plans to conduct additional checks.

Interested developers can apply to participate in testing at future OpenAI sessions at the following link.

RecentlyOpenAI adds improved image generation to GPT-4obut quickly removed it from access for free users.

.

Source: neowin