AI: Openai launches customizable models

OPENAI released, on Tuesday, two models of open generative artificial intelligence (IA), which can be personalized by the developers, after two years of debate on the ethical commitments of the Silicon Valley Star company. “We have created a model with open weight which reaches the performance of O4-Mini (Editor’s note: reduced version of GPT-4O, the most advanced model of Openai) and which can run on a high-end laptop (phew!), And another smaller on a phone,” said on X Sam Altman, the boss of the start-up. “Super proud of the team; A great technological victory, “he added.

A model with open weight is a model of AI whose parameters fixed during the training phase (“weight”) are made public, allowing developers to adjust them to their convenience. This does not mean that the model is entirely “open source” because the code and training data, in particular, can remain subject to restrictions.

Called GPT-OS-120B and GPT-OS-20B, the two newcomers provide the same security guarantees as the most advanced models in OpenAI, according to the company. “What is unique with these models is that people can use it directly on their own computer infrastructure, locally, with their own data, which may not have the right to leave their computers or servers in their country,” said the co-founder of Openai Greg Brockman during a press briefing.

OPENAI launched at the end of 2022 the wave of generative AI with Chatgpt. Since then, the tech giants like Google, and a handful of startups, compete with models and assistants that are ever more efficient, with a target of an AI called “general”, or “superintelligence”, with cognitive capacities superior to those of humans.

But at its foundation in 2015, Openai had indicated that it would provide the public research and information on its models. She had hardly done it so far, drawing criticisms of many supporters of this method, and one of the initial founders of the organization, Elon Musk.

“Since our beginnings, Openai’s mission has been to guarantee a general AI that benefits all of humanity,” said Sam Altman on Tuesday. “With this in mind, we are enthusiastic about the idea that the world is building on an open technological basis, created in the United States, based on democratic values, accessible for everyone and at the service of the greatest number,” he continued.

The American giant of Meta social networks, the Elon Musk Xai IA start-up and the French Mistral, in particular, publish models with open weights. Above all, the Chinese Deepseek released in January Deepseek R1, a model with open weight that took short tech environments thanks to its low -cost performance. “I think we were on the wrong side of history and that we have to find a different open source strategy,” said Sam Altman soon after, during a question and answer session.