Has suffered several delays, but finally OpenAI has launched its new Open weights language model That, he says, offers higher performance than others in his class in reasoning tasks. The company states that GPT-Ossthat’s what it is called, it is optimized so you can execute efficiently in consumer hardware.
GPT-Oss is available in two versions: one of 120,000 million parameters and another of 20,000 million. The first can work with a single GPU that has 80 GB of memory -That is, from the professional field- and offers a performance similar to that of the model O4-mini of Openai, while the lightest version yields like the O3-mini and need only 16 GB of memorymore accessible in the consumer market.
Both versions are already available for download on platforms such as Hugging Face, Databricks, Azure and Awslow Apache 2.0 licensewhich allows its modification and commercial use without restrictions. If you do not want to install it on your equipment, it can also be used through the browser, such as chatgpt.
What is a model of open weights
A open weights model It is an artificial intelligence model whose pesos -That is, the trained parameters that define their behavior- are published and can be discharged, analyzed, modified and reused freely, even for commercial purposes.
By contrast, a model of Black box as Chatgpt maintains its private weights and can only be used through a API controlled by the owner company.
The difference between an open weights model and another Open Source or of open source It is that the first allows access to the model trained, but not necessarily to the source code, while the second includes it complete.
Why now launches OpenAi GPT-Oss
GPT-Oss is Openai’s first open peso model in more than Six yearsbeing the last GPT-2three years before the launch of Chatgpt. Until recently, the CEO Sam Altman He had alleged security concerns as the main reason not to release models of this type. However, The rise of open models, which offer lower costs and greater customization capacity, has changed the panorama.
In January, after the irruption of DeepseekAltman acknowledged that Openai had been ‘On the wrong side of history’ By not launching open models.
Now, Openai seeks to recover land with a model of open weights that, according to the company, is capable of Reason, navigate online, write code and operate agents Through your current APIs.
‘Many people are surprised to know that the vast majority of our clients already use open models. We wanted to cover that hole and allow us to use our technology on all fronts‘, he explained Chris CookOpenai researcher on this launch.
Security and performance
In terms of security, Openai states that GPT-Oss is its more rigorously proven model to date and that it has been evaluated by specialized external companies to ensure that it does not involve risks in areas such as the Cybersecurity or the Biological weapons.
In addition, the model shows its Reasoning chain to facilitate the detection of undue behaviors, deception or improper uses. The information it generates is only in Text format And, like all OpenAi models, Your training data has not been revealed.
Openai has not published direct comparisons between GPT-Oss and other open models such as Calls (Meta) or Deepseek. However, both versions of the model offer a performance similar to that of closed reasoning models of OpenAI in programming tasks and tests such as Humanity’s last exam. ‘These models are incredible. The team has come out with this, ‘he said Greg BrockmanOpenai co -founder.
Although OpenAI has not committed to launching new versions of GPT-Oss regularly, he hopes that this model Be useful forsmaller developers and companies that seek to have more control over the use of their data. ‘We have always believed that, if you reduce access barriers, innovation shoots. Let people experience and do incredibly surprising things, ‘Brockman concludes.