OpenAI Just Released Its First Open-Weight Models Since GPT-2

OpenAI has excluded its first Open-Open models for more than five years. Two languages models, GPT -OS -120B and GPT -OS -20B, can operate locally on the customer device and subtle for certain purposes. For OpenThey represent the recent strategy of focusing on owned releases, since the company is more broad and more open, group of AI models that are available for users.

“We are interested in creating this model, the results of billions of dollars research, the world is available to get AI to most possible people in the world,” OpenAI CEO Sam Altman said in an emailed statement. GPT -OSS -120B and GPT -OS -20b both formally embraced face, is available for free download for a popular hosting platform for a popular hosting platform AIThe The latest Open-Ojan model was published by OpenAI GPT -2Back in 2019.

What separates the Open-Wojan model is its “weight” publicly available, which means that anyone can get an idea about how it processes it can make an idea about internal parameters. Instead of reducing the Opena -owned models with a free option, Cofowner Greg Brocman views this release as the “complement” of the services provided by the company, as is the current application programming interface used by many developers. “Open-Ojan models have a very different force,” Brookeman said in a briefing with journalists. Different from ChatzPT, you can run a GPT-OS model on the back of the Internet and on the back of the firewall.

Using both GPT-OS models Chain-off Approach, which Opena was first deployed to its O1 model The last fall. Instead of simply giving an output, the generator AI tools of this method pass through multiple steps to answer a prompt. These new text-cable models are not multimodal, but they can browse the web, to help cloud-based models work, execute code and navigate the software as a one. You have an agentThe Small of two models, GPT -SS -20b, with more than 16GB memory is compact enough to run locally on a customer device.

Available under OpenAI under two new models Apache 2.0 licenseA popular choice for Open-Ojan models. With Apache 2.0, models can be used for commercial purposes, re -distributed and include as part of other licensed software. OPEN-OPON MODER MODERS Published from Alibaba Queen As well as Mistral Manage under Apache 2.0 also.

In March, the publication of these open models was initially delayed for further security tests. Publishing an Open-Opon model is more dangerous than the Close-Off version because it removes obstacles around who can use the equipment and one can try to subtle a version of GPT-SS for involuntary purposes.

OpenAI usually customized the startup Open-Wojan option in addition to its owned models, so to see how it is possible to use the “bad actor” to see how it is possible to abuse it possible. “We have internalized the model internally in this risky area,” said Eric Wallace, an openinge protection researcher, “and measured how high we can make them. “In the Open-Opening Examination, Open-Ojan model did not reach high level of risk, as was measured by it Preparation structureThe