In August 2025, OpenAI took a strategic turn with the launch of GPT-OSS, a family of models that marks its return to the world of open source. In contrast, GPT-3, GPT-4, and GPT-5 have not been released open source. They are offered only through API or ChatGPT.
GPT-OSS is distributed with open weights under the Apache 2.0 license, which means that any developer or company can download, run and customize them on their own infrastructure.
This decision responds to a double commitment: promoting innovation in the technology community and offering greater control to those who want independence from the cloud.
Technological accessibility depends largely on the models not being closed boxes. The binomial between Software and Artificial Intelligence is strengthened when we opt for open models that allow the community to audit and improve the technology.
Intelligence without the limitations of a closed model
GPT-OSS is available in two sizes: 20B and 120B parameters. The latter is, to date, the most powerful open weights model released by OpenAI. Both use a Mixture-of-Experts (MoE) architecture, designed to optimize inference efficiency and speed.
Thanks to this, they can handle complex reasoning, execute tools and function as autonomous agents in different applications.
The impact of GPT-OSS goes beyond the technical. For small companies, universities or research laboratories, having models of this level without the recurring costs of the API represents a unique opportunity.
It allows them to experiment, build their own products and advance artificial intelligence projects without the access limitations that characterize closed models.
This post is also available in: