usatoday24
Although his name suggests a certain vocation of technological opening, Openai had been moving away from that spirit for some time. From long before the launch of Chatgptin November 2022, the company did not publish an open weight model. In fact, it is the first open weight model since 2019 (GPT-2).
That has just changed.
This Tuesday, OpenAi has announced GPT-Ossan open weight model under Apache 2.0 license that can be downloaded for free and, in its small variant, run even in a laptop. A movement that nobody expected at all, especially since all eyes were put in Lat the imminent arrival of GPT-5.
What is GPT-Oss? GPT-Oss is a language model developed by OpenAI, capable of understanding text instructions and generating answers with a high level of coherence, context and utility. It is a variant of the GPT family, the same technology that gave rise to Chatgpt, but with a key difference: its weights are open. That is, you can download it, run it on your own computer and adapt it to your needs.
For the first time in years, OpenAI launches a powerful model that is not tied to its platform. You can use it without internet connection, without depending on your servers and freely to customize it. And besides, it is free. That opens the door for developers, small companies or even advanced users can experiment with AI in their own terms.
What means that the weights are open? When we talk about the “pesos” of a model are open, we refer to the internal numerical values that you have learned during your training. They are, in essence, the knowledge of the model: what allows you to reason, write, translate or generate useful answers.
That OpenAi has released those weights means that anyone can download them, analyze them and use them outside their servers. It is not the same as releasing the complete source code or the data with which it was trained, but it is an important step: it allows you to execute the model in own environments, adapt it to concrete tasks or integrate it into products without having to go through the Openai API.
So is it a really “open” model? Now, the weights are open does not mean that the model is completely open. GPT -Oss is not Open Source strictly speaking: OpenAi has not published the data with which the model or the full code of the training process trained.
What has released, as we say, is the final result, the parameters learned, under a permissive license (Apache 2.0). That is why we talk about an open weight model, not open source, According to the standards of organizations such as the Open Source Initiative (OSI).
Can I use it on my computer? It depends on the model. GPT-Oss occurs in two versions: one of 120,000 million parameters and a lighter of 20,000 million. The most powerful version requires a GPU of at least 80 GB of memory (professional environments). The 20B version is much more accessible: it can be executed on computers with 16 GB of memory, which opens the door to use in personal portables and PCs. It is already compatible with tools such as LM Studio (which we have used to try Deepseek and other models)flame.CPP or VLLM, which allows you to try it at home without connecting to the cloud or depending on OpenAi. Of course, it is recommended to have some technical experience to put everything in motion.
It is compatible with tools such as LM Studio, Llama.CPP or VLLM.
Where do I download it? Is it really free? Weights are available for free download mainly in Hugging Faceand can also be deployed through platforms such as Azure, AWS or Databricks. The two variants, 120b and 20b, together with the necessary documentation are available.
OpenAI publishes GPT-Oss under Apache 2.0 license, which means that you can use it, modify it and adapt it for commercial purposes. Do not pay for download or use it at home. If you choose to execute it in the cloud or on a payment infrastructure, the cost will depend on the supplier you use.
Is it easy to use or do you have to have a certain level of knowledge? Using GPT-Oss is not as complicated as training a model from scratch, but it is not a solution that works with two clicks. The models include clear instructions and are integrated into platforms such as LM Studio or Llama.cpp, which allow them to be executed at local ease.
That said, you will need to understand basic concepts such as execution environment, models loading and parameter configuration. If you have already worked with other open weight models, it will not cost you. If not, there are guides available, but you will have to devote some time to put it to function as it is.
What can this model do? GPT-Oss is a multi-support model. You can reason, generate text, write code, answer questions, search for information and operate external tools, such as python browsers or interpreters. All this is also compatible with agents based on workflows, useful for more complex and automated tasks.
Openai states that the 120B model achieves a proximity with O4-MINI in reasoning tests, while the 20B model offers similar results to O3-MINI in Common Benchmarks, with a good performance in programming, reasoning for steps and use of tools. It is not the most powerful model of the house, but one of the most versatile that can be used without paying or depending on your API.
Images | OpenAI
In Xataka | A watchmaking pump threatens the most powerful alliance of technology: Microsoft’s AGI clause
Images | OpenAI
In Xataka | A watchmaking pump threatens the most powerful alliance of technology: Microsoft’s AGI clause
GIPHY App Key not set. Please check settings