Openai enters an open source AI race with new thinking models-while guarding its IP address

“We are excited to make this model, as a result of billions of dollars in research, available to the world to obtain Amnesty International in the hands of most people as possible,” said Sam Altman, CEO of Openai. “As part of this, we hope that this version will enable new types of research and create new types of products.” He stressed that he was “enthusiastic about the world to build on the Amnesty International Open Phalae that was created in the United States, based on democratic values, available for free for all and wide interest.”

The CEO of Openai, Sam -German, disturbed the upcoming models in March, two months after confessing, following the success of the open models in China from Depsic, that the company was.On the wrong side of history“When it comes to opening their models for developers and builders. But while weights are now general, experts notice that the new Openai models are barely“ open. ”There is no way in any way to give up the crown jewels: ownership structure, guidance mechanisms, training data and methods that operate their most advanced models-including the GPT-5 that long Waiting, it is expected to be widely released at one time this month-by providing under the winding.

Openai targets artificial intelligence builders and developers

The name of the two new models-GPT -SS-120B and GPT-SSS-20B-may not be not subject to non-engineers, but this is because Openai puts its attention on the AI’s builders and developers who seek to build quickly on cases of use in the real world on their own systems. The company noticed that the two largest models could work on one Nafidia 80 GB chip, while the smallest is suitable for consumer devices like a Mac laptop.

Greg Brockman, co -founder and president of Openai, admitted an invitation to the prior press that “has been a long time” since the company issued an open model. He added that “something we consider complementary to the other products that we launch,” in addition to the royal Openai models, “the combination of speeding our mission of ensuring that the API programming interface benefits from humanity.”

Openai said that new models are well working on thinking criteria, which appeared as major measurements to perform artificial intelligence, with models of Openai, anthropology, Google Deepseek are strongly competing for their ability to treat multi -step logic, generate code, and solve complex problems. Since the open source of Deepsek R1 rocked the industry in January with its logical capabilities at a much lower cost, many other Chinese models followed their example – including QWEN models and Moonshot Ai from Alibaba. While Openai said in the prior press, the new new open -weight models are a pre -emptive effort to provide what users want, as it is also a strategic response to intensify the competition of open sources.

It is worth noting that Openai refused to determine its new open-weight models against open source Chinese systems such as Deepseek or QWEN-despite the fact that these models recently outperformed the competitors of the United States over the main thinking standards. At the press conference, the company said it is confident in its standards against its own models and that it will leave it to others in the artificial intelligence community to test more and “form their minds.”

Avoid intellectual property leakage

The new OpenAI models are designed with an open weight using the structure of experience (MEE), where the system is activated only “experts” or sub -networks, it needs specific inputs, instead of using the entire form for each query. Dylan Patel, founder of the Research Company in half of the analysis, indicated in a publication in x Before the version that trained Openai on models only using the well-known components of the audience for architecture-however, the building blocks that he used already used in the open source community. He stressed that this was a deliberate choice – by avoiding any special training techniques or architecture innovations, Openai can issue a really useful model without leaking any intellectual property that already works to operate its border models such as GPT 4O.

For example, in a model card accompanying the release, Openai emphasized that the models use a mixture of expert structure (MEE) with 12 active experts out of 64, but it does not describe the guidance mechanism, which is a decisive and owned part of the structure.

“You want to reduce risks to your work, but you (also) want to be extremely useful to the public,” Alexa Gordic, former Google Deepmind A researcher, he told Fortune, adding that companies are like Dead Mistral, which also focused on open weight models, also did not include royal information.

He said: “They reduce the IP leakage and remove any danger to their basic works, while at the same time they share a useful artifact that will enable the ecosystem and developers to start operating.” “It is by defining the best they can do in view of these two opposing goals.”