Published by NewsPR Today | August 2025
AI research firm OpenAI has released two new artificial intelligence models that can be downloaded and run directly on personal computers, with no internet connection required.
The new models, named gpt-oss-20b and gpt-oss-120b, are the company’s first “open-weight” models made available to the public since the release of GPT-2 several years ago. This means anyone with a powerful enough computer can run the AI locally, see how it processes information step-by-step, and even modify it.
OpenAI has released two versions: a smaller, more compact model designed for modern laptops and desktops, and a much larger, more powerful version intended for high-end workstations and data centers. While users are free to build on the models, OpenAI confirmed that safety and censorship guardrails remain in place.
So, what kind of computer do you need to bring this technology home?
Hardware Needed for the Everyday Model: gpt-oss-20b
The smaller model, gpt-oss-20b, is designed for consumer devices. According to OpenAI, the main requirement is 16GB of RAM.
This memory can be your computer’s main system RAM, the dedicated memory on your graphics card (VRAM), or a combination of both. While 16GB is the minimum, performance will be smoother and faster with more memory.
For example, users have reported running the model smoothly on a MacBook Pro M3 with 18GB of memory. AMD suggests pairing its new Ryzen AI 300 series processors with 32GB of RAM for an ideal experience. The key takeaway is that a modern computer with at least 16GB of memory should be able to run it.
Hardware Needed for the Powerhouse Model: gpt-oss-120b
The larger model, gpt-oss-120b, is a different beast entirely, demanding a massive 80GB of available RAM.
This model is primarily built for professional data center graphics cards that have 80GB of VRAM or more. However, it is possible to run it on a high-end home workstation by combining system memory with graphics card memory. For instance, a PC with 64GB of system RAM and a 24GB graphics card (like an NVIDIA RTX 3090 or 4090) would meet the requirement.
This puts the 120b model firmly in the territory of seriously powerful—and expensive—workstation PCs.
How to Download and Run the Models
For those who want to try them, the models are available on the software repository Hugging Face. The easiest way for most people to run them is with a free tool called Ollama, which works on Windows, Mac, and Linux.
After installing Ollama, you can run the model of your choice by opening a command terminal and typing:
- For the smaller model: ollama run gpt-oss:20b
- For the larger model: ollama run gpt-oss:120b
Other tools like LM Studio and the new (still in preview) Windows AI Foundry also support the models.
Can It Really Run on a Smartphone?
While OpenAI CEO Sam Altman suggested the smaller model could work on smartphones, that appears to be an ambitious claim for now.
While technically possible on a future phone with 16 GB of RAM, current reports indicate the technology is more suited for new AI-powered laptops, like the upcoming Copilot+ PCs. For now, running this level of AI on today’s phones is not a realistic experience. However, this release signals a clear future where powerful AI will eventually run on the devices in our pockets.
