“Something big, but small.” It is with these words that Sam Altman, the founder of Openai, announced the launch this Tuesday, August 5, two new open source models. A first from GPT-2, in 2019.
A change of strategy
Called “GPT-OS-120B” and “GPT-OS-20B”, these two new models of open source language are published under the Apache license. Therefore, the implementation, organizations and individuals can consult and modify the programming code for these models for free to create their own tools and customize them in their own infrastructure.
When it started in 2015, OpenAI had adopted an open source approach. Your mission? Develop general artificial intelligence for the good of humanity. But over the years, the new company that caused Chatgpt had finally changed their minds to strong competition in the sector. In 2023, the company even indicated that this approach was not correct for safety and competition reasons.
But in 2025, after the success of the Chinese AI models as Deepseek, Sam Altman made a back machine again. The boss acknowledged that Optai had found himself “on the wrong side of the story” by not publishing his own open models. Expected for several months, the publication of these models had postponed at the beginning of summer.
“We need time to perform additional security tests and examine high -risk areas,” said Sam Altman last July. Now it is made.
A powerful model on the smartphone
Therefore, the GPT-OS model is available in two versions: 120 and 20 billion parameters. The most powerful version can operate in a single NVIDIA GPU and offers a performance similar to that of the existing OPENAI O4-mini model. For its part, the model with 20 billion configurations can be executed on a computer, or even on a smartphone with at least 16 GB of memory. It is equal to the performance of the O3-mini.
Whatever the version, the two GPT-Oss allow you to perform reasoning tasks, navigate the web, write code and use agents through the existing company’s API.
On the security side, Openai says that GPT-OS is its most rigorously proven model to date. According to documents published by the company, it was tested by external security companies to guarantee their lack of risks in areas such as cybersecurity and biological weapons.
And the launch of Operai open source models should not be the only announcement of the week. In X, the boss evokes “an even greater novelty later in the week.”
Several experts evoke the next GPT-5 launch, their long-awaited new language model. Sam Altman announced a first overview on Monday, August 4 on the social network, in the turn of a conversation in the Pantheon television series.
Source: BFM TV
