Open AI CEO Sam Altman speaks on the annual Snowflake Summit in San Francisco, California on June 02, 2025.
Justin Sullivan | Getty Pictures Information | Getty Pictures
OpenAI on Tuesday launched two open-weight language fashions for the primary time because it rolled out GPT-2 in 2019.
The text-only fashions are referred to as gpt-oss-120b and gpt-oss-20b, and are designed to function lower-cost choices that developers, researchers and firms can simply run and customise, OpenAI mentioned.
An artificial intelligence mannequin is taken into account open weight if its parameters, or the weather that enhance its outputs and predictions throughout coaching, are publicly accessible. Open-weight fashions can provide transparency and management, however they’re completely different from open-source fashions, whose full supply code turns into accessible for individuals to make use of and modify.
A number of different tech firms, together with Meta, Microsoft-backed Mistral AI and the Chinese language startup DeepSeek, have additionally launched open-weight fashions lately.
“It has been thrilling to see an ecosystem develop, and we’re excited to contribute to that and actually push the frontier after which see what occurs from there,” OpenAI President Greg Brockman advised reporters throughout a briefing.
The corporate collaborated with Nvidia, Advanced Micro Devices, Cerebras, and Groq to make sure the fashions will work properly on a wide range of chips.
“OpenAI confirmed the world what might be constructed on Nvidia AI — and now they’re advancing innovation in open-source software program,” Nvidia CEO Jensen Huang mentioned in an announcement.
The discharge of OpenAI’s open weight fashions has been extremely anticipated, partially as a result of the corporate repeatedly delayed the launch.
In a post on X in July, OpenAI CEO Sam Altman mentioned the corporate wanted extra time to “run extra security assessments and overview high-risk areas.” That got here after a separate post weeks earlier, the place Altman mentioned the fashions wouldn’t be launched in June.
OpenAI mentioned Tuesday that it carried out in depth security coaching and testing on its open-weight fashions.
It filtered out dangerous chemical, organic, radiological and nuclear information throughout pre-training, and it mimicked how dangerous actors might attempt to fine-tune the fashions for malicious functions. By way of this testing, OpenAI mentioned it decided that maliciously fine-tuned fashions weren’t in a position to attain the “excessive functionality” threshold in its Preparedness Framework, which is its technique for measuring and defending towards hurt.
The corporate additionally labored with three unbiased skilled teams who supplied suggestions on its malicious fine-tuning analysis, OpenAI mentioned.
OpenAI mentioned individuals can obtain the weights for gpt-oss-120b and gpt-oss-20b on platforms like Hugging Face and GitHub underneath an Apache 2.0 license. The fashions might be accessible to run on PCs by means of packages corresponding to LM Studio and Ollama. Cloud suppliers Amazon, Baseten and Microsoft are additionally making the fashions accessible.
Each fashions can deal with superior reasoning, instrument use and chain‑of‑thought processing, and are designed to run anyplace — from shopper {hardware} to the cloud to on-device functions.
Customers can run gpt-oss-20b on a laptop computer, for example, and use it as a private assistant that may search by means of information and write, OpenAI mentioned.
“We’re excited to make this mannequin, the results of billions of {dollars} of analysis, accessible to the world to get AI into the palms of the most individuals doable,” Altman mentioned in an announcement Tuesday.
–CNBC’s Jordan Novet contributed to this report
