Despite the company’s name, Open has not dropped the open version of its AI models since GPT2 in 2019. It changed on Tuesday, as CEO Sam Limitman shared two new Open Whites, argument AI models, called GPTOS -1220B (120 billion parameters) and GPT -SOS 20B (20 billion parameters).
If Open Whites is a new piece of Ai Jirgun for you, don’t worry. In the easiest possible terms, open Whites is a category of AI models that have electric products such as chat boats, image and video generators. But they are philosophically different from the technology you can now use. Chat GPT, Gemini and Co -Pilot are powered by all closed models, which means we have no real insights about how these black box machines work. Open Whites Models allow us to peek at the mechanical wizard behind the scenes, so talk.
GPT-SOS is a big deal. This is a sophisticated open-weight reasoning model, compared to O4-mini, with which you can locally run on your computer (or phone with small size) locally. We believe this… I am the best and highly usable open model.
– Sam Altman (@SAMA) August 5, 2025
You do not need to be developer or machine learning expert to understand how these open models work or even run them themselves. You need everything to learn about open weight and open source AI models.
What is an open weight AI model?
All AI models weigh, which are features or elements. Models are trained so that some contacts can get overweight, or cost.
An open-weight model does exactly what his name shows-Azan is publicly available, as stated by the Federal Trade Commission. Developers can see these weights and how they are used in the creation of AI models.
CNET told CNET, “The most valuable thing in the argument, big (language) models is actually weight. If you have weight, you can do a lot, which is something different from traditional software,” said CNET, Assistant Professor of Computer Science in MIT and its computer science and artificial intelligence lab (CSAIL).
For example, in a sentence, a chat boot is made to predict the next logical word for the next log. It is trained to collect words in its results, which may often appear together in the training data, perhaps in a logical order. Words that appear more often than one another can be given overweight than words that are often not manifest with each other.
These weights are only numbers, but open -weight models also come with the map.
“In open weight (models), you get weights, which are these numbers, and you find these weights in the nerve network structure, so the nerve network layers, in fact, to be able to operate it.
Open Whites models are primarily aimed at developers, which can integrate the model into existing projects, such as helping to build AI agents. For “determined hobbies”, as Khatab said, you can use spears to run the model locally on your laptop, which can help address incoming privacy concerns, for example, using AI through a company’s mobile app. Researchers will also see how AI works internally.
How do Openai’s new Open Weights model steak up?
The new Open Whites Model comes in two sizes, 120 billion parameters (128 experts and 128K context window) and 20 billion parameters (32 experts but a single 128K context window). Experts refer to the number of sub -nerve networks of a model, and the context of the context explains how much information the model can process and add it in response. A large number of for both indicate that a model is worth more sophisticated and has more firepower.
In terms of performance, the Openi is reporting that the 120B model “got the closest paratide” with its latest reasoning model, O4-Mini, on a single 80 Gighek GPU benchmark on the Benchmark. The 20B Open Whites model similarly performed with O3-Mini and run on a 16-gigabyte device-which means that this small open-white model can be largely run on laptops and some smartphones. (As all AI models run locally, your speed will depend on the firepower of your device.)
The model will be available under Apache 2.0 license, which is a type of open source friendly license. You can check the safety training models in cards and paper more deeply in the paper, get the points from the open developer’s guidelines, and now see the weight for yourself on hugs and gut hubs.
CNET’s Emad Khan installed a GPT-OSS-20B on a PC, which was specifically designed to run a local AI model.
Is Open Weight Like Open Source AI?
Open Whites model is related to the open source AI, but they are not the same. As a concept, open source refers to the software that has no proprietary owner, whose source code is publicly available and can mostly be used under open source license. Open Source Initiative, a non -profit that advocates open source software, explains Open Source AI “a system that is created under conditions that give users the freedom to use, study, modify and share it.”
Openweights AI model is not like an open source model. One way to think about the difference between the two is akin to baking, Saba Vasodone, the chief operating officer of Mozilla dot or the senior vice president in Mozilla Corporation, told CNET.
“There is an open -weight model giving you this baked cake and says, ‘Oh, it’s made of flour, sugar and eggs.’ Vasodivan said, “They are weight, but if anyone gives you full guidance, and how many instructions and its definite details are in it, then this is open source.”
Open weight models, the things that were not disclosed are data on which the model was trained and the code was used to train it. Training data is a point of conflict between AI companies and humans. AI companies are high quality, high quality, human -made materials to improve and improve their models. Some companies collect these data through licensing contracts, but some publishers and creators have filed a lawsuit, alleging that AI companies are illegally receiving their rights content. –
Training data, regardless of its origin, is one of the most valuable things AI company. But it is likely that it will not be included in the release of any openweights model.
“I think responsibility is due to a sheer scale (data), because many of them are licensed and you should not share it, I think it is appropriate to understand that at least no company will soon release it soon, maybe, perhaps ever,”.
Vasodion said that the open source AI really comes with more publicly available information. Without additional information, open weight models can be difficult to re -train or prejudice. “You are still quite substantially in the dark about how it was built or its data,” said Wasudwan.
See this: How do you talk about Chat GPT matters? Why is here
04:12
Why should I care about Open Source AI?
These differences between open source, open Whites and closed models will potentially not affect your average experience using AI chat boot. But they are important features for developers to choose and use some AI models, and they are more widely important to us to understand this technology that is entering our online lives.
Khatab said he had no guarantee that his open -weight models would reflect the internal works of what was within its closed models. But for a company whose product is equivalent to General AI, any light shed on its working method is to make sure that people who use it and those who design it behind the screen. When people begin to enter mournful clothing with new models, we will find more information and see how it is moving on the industry.
Overall, the main philosophy of open source is that technology improves when more and more people can access it, such as educators who can study it and security experts who can expose weaknesses. “You have to allow people to build, scale and innovate and throw holes and improve holes,” said Vasodivan.
For more, check our initial guide for Chat GPT and how AI is changing our search experience.
