On running GPT-OSS-20B on my Macbook Pro

I’m been running gpt-oss-20b on my Macbook Pro for some time now. I typically run different LLMs on my Macbook using ollama as the basis for this. While I’ve tried various models, this model from OpenAI is the one I found that gave me the best answers to the questions I’ve been giving it. I even prefer the answers I get with gpt-oss-20b on my Mac over those I get from larger SaaS models provided by OpenAI, Anthropic and Google.

I’ve found it very good when it comes to asking it questions concerning lifestyle changes. For instance, I asked it for:

  • ideas on how to celebrate a big birthday
  • a 12 month plan to do Swedish death cleaning on my house
  • a 12 month plan to get my estate in order
  • a 12 month plan to get back in shape after a long layoff

I’ve found it not very good at all with questions around food and music:

  • coming up with music playlists based on certain criteria (over a third of the songs were hallucinations!)
  • coming up with spice mixes for cooking meats (the proportions were way off compared to tested recipes)

Like all LLMs, I found it works well as a means to generate ideas. It works ok or poorly when you start expecting exactitude in the answers. YMMV.

Other things I like about gpt-oss-20b running on ollama on my Macbook:

  • I can see it go through it’s reasoning when it comes up with the answer, which is a nice feature.
  • For some reason the default output format is one to my liking, and I can easily move it into Microsoft Word or Apple Pages for further formatting
  • I especially like it because I can ask personal questions and not have those being captured by a service in the cloud. It provides an extra layer of privacy.

I am running gpt-oss-20b and not gpt-oss-120b due to the capacity of my Macbook Pro. Indeed, even gpt-oss-20b takes up a fair amount of resources on my Macbook, despite it having a GPU to do the heavy lifting. I am not in a rush for these answers though, so it is fine. I’d run 120b if I could, but I have been happy with the response with 20b.

If you have a machine with a GPU and enough RAM and disk space, consider downloading ollama and this model and give it a go. Let me know what you discover!