As you may have seen, Openai has just published two new models of AI – GPT -OS -20B and GPT – OSS -120B – which are the first open -weight models from the company from GPT – 2.
These two models – one are more compact, and the other much larger – are defined by the fact that you can execute them locally. They will work on your desktop computer or laptop – directly on the device, without having to go online or press the power of the cloud, provided that your equipment is powerful.
So, you can download either the 20B version – or, if your PC is a powerful machine, the 120B spin – and play with it on your computer, check how it works (text with text to text) and how the model thinks (all its reasoning process is divided into steps). And indeed, you can modify and build on these open models, although the safety railings and censorship measures will of course be in place.
But what type of equipment do you need to execute these AI models? In this article, I examine the requirements of PC specifications for GPT-OS-20B-The smaller model of 21 billion parameters-and GPT-OS-15B, which offers 117 billion parameters. The latter is designed for using the data center, but it will work on a high-end PC, while GPT-OS-20B is the model designed specifically for consumer devices.
Indeed, when they announce these new AI models, Sam Altman referred to 20B working not only on ordinary laptops, but also smartphones – but it is enough to say that it is a ambitious claim, to which I will come back later.
These models can be downloaded from Hugging Face (here is GPT-OS-20B and here is GPT-OSS-120B) under the APACHE 2.0 license, or for simply, there is an online demo that you can consult (no necessary download).
The smallest GPT-OS-20B model
Minimum RAM required: 16 GB
Openai’s official documentation simply has a required amount of RAM for these AI models, which in the case of this GPT-AS-20B Plus compact effort is 16 GB.
This means that you can execute GPT-OS-20B on any laptop or PC which has 16 GB of system memory (or 16 GB of video RAM, or a combo of the two). However, this is really one more case, the happiest – or faster. The model can struggle with this minimum of 16 GB, and ideally, you will want a little more.
With regard to CPUs, AMD recommends the use of a Ryzen AI 300 series processor associated with 32 GB of memory (and half of this, 16 GB, defined on variable graphic memory). For the GPU, AMD recommends any RX 7000 or 9000 model which has 16 GB of memory – but these are not difficult and fast requirements as such.
Really, the key factor is to simply have enough memory – the 16 GB allowance mentioned, and preferably have all of this on your GPU. This allows all the work to take place on the graphics card, without being slowed down by having to unload part of the PC system memory. Although the so-called mixture of experts, or MOE, the OPENAI design has used here helps to minimize such a trail of performance, fortunately.
Anecdotal, to choose an example drawn in Reddit, GPT-OS-20B works well on a MacBook Pro M3 with 18 GB.
The largest GPT-OS-120B model
RAM required: 80 GB
This is the same global agreement with the GPT-AUS-20B model more strong, except as you can guess, you need a lot No more memory. Officially, that means 80 GB, although don’t forget that you don’t need to have this whole RAM on your graphics card. That said, this large AI model is really designed for the use of the data center on a GPU with 80 GB of memory on board.
However, the RAM allowance can be divided. Thus, you can execute GPT-AS-15B on a computer with 64 GB of system memory and a 24 GB graphics card (a NVIDIA RTX 3090 TI, for example, according to this Redditor), which makes a total of 88 GB of RAM grouped.
AMD’s recommendation in this case, at CPU, concerns its Ryzen AI Max + 395 high -end processor coupled with 128 GB of system RAM (and 96 GB of that allocated as variable graphic memory).
In other words, you are considering a very high-end laptop or desktop computer (perhaps with several GPUs) for GPT-AS-15B. However, you may be able to get out of it with a little less than the 80 GB of stipulated memory, including certain anecdotal reports – although I would not put it in any case.
How to execute these models on your PC
Assuming that you meet the requirements of the above system, you can execute one of these new GPT-OS versions on Olllama, which is the OpenAi choice platform to use these models.
Head here to enter Oilama for your PC (Windows, Mac or Linux)-Click the button to download the executable, and once the download is finished, double-click the executable file to run it and click on Install.
Then run the following two commands in Olllama to get, then run the model you want. In the example below, we execute GPT-AS-20B, but if you want the larger model, simply replace 20b with 120b.
ollama pull gpt-oss:20b
ollama run gpt-oss:20bIf you prefer another option rather than Olllama, you can use LM Studio instead, using the next command. Again, you can change 20b for 120b, or vice versa, as the case may be:
lms get openai/gpt-oss-20bWindows 11 users (or 10) can exercise the Windows AI foundry option (hat tip for the rod).
In this case, you will have to install Foundry Local – there is a warning here, however, and it is that it is always in preview – consult this guide for complete instructions on what to do. Also note that at present, you will need an Nvidia graphics card with 16 GB of VRAM on board (although other GPUs, such as AMD Radeon models, are finally supported – Remember, it is always a preview version).
In addition, the support of macos “arrives soon”, we are told.
What about smartphones?
As indicated at the start, while Sam Altman said that the smaller model operating on a phone, this declaration pushes it.
It is true, Qualcomm published a press release (as identified by Android Authority) on GPT -OS -20B operating on aircraft with a snapdragon chip, but it is more laptops – Copilot + PC which have Snapdragon X Silicon – rather than on smartphone processors.
The execution of the GPT-OS-20B is not a realistic proposal for today’s phones, although this can be possible in a technical direction (assuming that your phone has 16 GB + RAM). Despite this, I doubt that the results are impressive.
However, we do not move away these types of models that work properly on mobiles, and it will surely be in cards for an almost enough future.
