Same question, two routes, below. One leaves your computer. One never does.
On the left, a model behind an [API](/what-is-an-api-2026) on the provider's servers. On the right, a local model running on your own machine.
API vs local, side by side
Who sees your prompts
- API (provider's servers)
- The provider, under their terms
- Local (your machine)
- No one else, when run locally
How you pay
- API (provider's servers)
- Per token, every request
- Local (your machine)
- Hardware up front, electricity after
Which models
- API (provider's servers)
- Closed models (e.g. Claude, OpenAI's GPT)
- Local (your machine)
- Open-weight models (e.g. gpt-oss, Qwen)
What limits you
- API (provider's servers)
- Rate limits and budget
- Local (your machine)
- Your GPU's memory (VRAM)
Privacy is the big one. Ollama's FAQ: "Ollama runs locally. We don't see your prompts or data when you run locally."
The limit is memory. A model has to fit in your graphics card's VRAM to run well. Our VRAM calculator shows what fits your card.
Cost flips. An API costs per request; local costs up front. Cost to run compares the two for real models, and cloud GPU prices sit in between.
It's the same idea underneath: tokens, a context window, sampling. Only where it runs changes.
Try it with the rules off. In the sandbox, send a prompt containing a fake customer email both ways, then check where it went. Then pick a model too big for your card and see what happens.
Check yourself
0/4 got itSaved on this device only. No account, no streaks.
