With the recent release of the gpt-oss-20b model by OpenAI, we can finally try out one of its mid-sized language models directly on our own Mac, without relying on the cloud or any external connections. Unlike ChatGPT (available via browser or with a ChatGPT Plus subscription), which works over the web and needs an internet connection, this model can run entirely offline. Here’s how.
What is gpt-oss-20b and what do we need to run it
OpenAI has released two versions of this model: gpt-oss-20b and gpt-oss-120b. The difference between them lies mainly in the size of the data and the accuracy of their responses.
The first one, the smaller model, is the one we can run locally, at home. It’s designed for computers with a decent amount of power, like Macs with Apple Silicon chips.
It’s a fast and lightweight model, but it can also “make things up” at times. We should keep in mind that a local model like this has no access to the web or any additional tools. It only works with the data it has already been trained on. Still, it’s a great starting point for experimenting, learning, or simply seeing how these AIs perform when they don’t rely on internet access.
As for disk space, the gpt-oss-20b model requires around 15 GB of free space. And although OpenAI says 16 GB of RAM is enough, that’s really the bare minimum. The more memory we have, the smoother the experience will be. On a MacBook Air M4 with 16 GB, for example, asking it to say “hello” might take more than five minutes. And anything slightly more complex… quite a bit longer.
How to install and run the gpt-oss-20b model on a Mac
Installing the gpt-oss-20b model is extremely simple.
First, we need to download and install “Ollama”, a free app that serves as the gateway to any models we want to run locally. We can find it at ollama.com/download and download the Mac version directly.
Once installed, we open Terminal and type the following command:
ollama pull gpt-oss:20bThis command will download the gpt-oss-20b model onto our machine. Once we have it, we can launch it with:
ollama run gpt-oss:20bAnd that’s it. If we open the Ollama app, we’ll be able to select the model from its panel and start interacting with it by entering any prompt.
From the app settings, we can also enable “airplane mode”, a feature that blocks any external connections to ensure everything the model does happens entirely within the Mac. No cloud, no accounts, no logging.
And when we want to delete it? We can do so from Terminal with this simple command:
ollama rm gpt-oss:20bThe gpt-oss-20b model is a powerful and impressive tool. But since it runs locally, it needs to use all the resources it can access, which may slow down the Mac while the model is processing. Is it a good option for quick and reliable tasks? Not really. But it is a great opportunity to experiment with AI in an autonomous and private way.
On Hanaringo | If you haven’t updated to iOS 18.6 yet, you should do it as soon as possible