OpenAI is releasing a brand new open-weight mannequin dubbed GPT-OSS that may be downloaded without spending a dime, be custom-made, and even run on a laptop computer.
The mannequin is available in two variants: 120-billion-parameter and 20-billion-parameter variations. The larger model can run on a single Nvidia GPU and performs equally to OpenAI’s present o4-mini mannequin, whereas the smaller model performs equally to o3-mini and runs on simply 16GB of reminiscence. Each mannequin variations are being launched in the present day by way of platforms like Hugging Face, Databricks, Azure, and AWS below the Apache 2.0 license, which permits them to be extensively modified for industrial functions.
That is OpenAI’s first open-weight mannequin in over six years, years earlier than the debut of ChatGPT. Till earlier this 12 months, CEO Sam Altman cited security considerations as the principle motive for not releasing a follow-up. In the meantime, builders have flocked to open fashions resulting from their decrease price and customizability. In January, after the rise of DeepSeek, Altman said that OpenAI had “been on the improper aspect of historical past” by not releasing its personal open fashions.
Now, OpenAI is reasserting itself with an open-weight mannequin that it says can carry out reasoning duties, browse the online, write code, and function brokers by way of the corporate’s present APIs. “I feel lots of people are literally shocked to know that the overwhelming majority of our prospects are already utilizing numerous open fashions,” Chris Prepare dinner, an OpenAI researcher, stated throughout a media briefing. “We needed to plug that hole and permit them to make use of our expertise throughout the board.”
On the protection entrance, OpenAI says that GPT-OSS is its most rigorously examined mannequin thus far, and that it was examined with exterior security corporations to make sure it doesn’t pose dangers in areas like cybersecurity and organic weapons. The mannequin’s chain of thought, or seen course of used to reach at a solution, is proven “to watch mannequin misbehavior, deception and misuse,” in keeping with an organization press launch. Its output is text-only and, like all of OpenAI’s fashions, GPT-OSS’s coaching information is undisclosed.
“The group actually cooked with this one.”
OpenAI hasn’t shared benchmarks evaluating GPT-OSS to different open fashions like Llama, DeepSeek, or Google’s Gemma. Each variants of GPT-OSS carry out equally to OpenAI’s closed reasoning fashions on coding duties and checks like Humanity’s Final Examination. “These are unimaginable fashions,” stated OpenAI cofounder Greg Brockman. “The group actually cooked with this one.”
OpenAI isn’t committing to a launch schedule for future variations of GPT-OSS, however it hopes that the mannequin might be utilized by smaller builders and firms that need extra management over how their information is used. “We’ve all the time believed that for those who decrease the barrier to entry, then innovation simply goes up,” stated Brockman. “You let folks hack, then they are going to do issues which are extremely shocking.”