How to import your own model into a Mac AI app
- You are not limited to the tiers we ship — you can point Outlier at model weights you already have.
- Imported models get the same surface as built-in ones: chat, tools, and the local API server.
- The app checks the model's chat template on import, because a model prompted in the wrong format produces confident nonsense.
- Everything stays on disk. Importing a model is a file operation, not an upload.
Every local-AI app has an opinion about which models you should run. That is useful right up until you have your own — something you fine-tuned, something niche, or simply a build you trust. Then the opinion is in the way.
Why bring your own model
Three reasons keep coming up. You fine-tuned something on your own data and want to actually use it. You need a model specialized for a domain no general tier covers. Or you have a specific build you have already validated and do not want silently swapped.
In all three the model is the easy part — what you actually want is the application around it: a real chat UI, tool calling, project context, and an API endpoint.
Importing one
- Have the weights on disk in a supported format.
- Add it in Settings as a custom model.
- Let the chat-template check run. This matters more than it sounds — see below.
- Pick it like any other tier from the model switcher.
The chat-template trap
A model expects its prompt wrapped in a particular way. Feed a model the wrong wrapper and it does not error — it keeps generating, often continuing your prompt as a document instead of answering it. The output looks like a quality problem when it is really a formatting problem.
We have been bitten by exactly this: a tier prompted in another family's chat format echoed its input back, and it read as the model being broken. So the app inspects the imported model's own template rather than assuming a family, and tells you when something looks wrong instead of quietly producing nonsense.
What to expect
- RAM is the constraint. A model has to fit in unified memory alongside everything else you're running.
- Tool calling depends on the model. Not every set of weights can drive an agent loop, regardless of the wrapper around it.
- Quality is yours to own. We publish measured numbers for our tiers; an imported model comes with whatever its author measured.
Common questions
Can I run my own fine-tuned model on a Mac?
Yes. Import the weights as a custom model and it appears in the switcher alongside the built-in tiers, with the same chat, tools and local API available to it.
Do imported models work with the local API server?
Yes. Once imported, a model is selectable like any tier, including for the OpenAI-compatible endpoint.
Why does my imported model reply with nonsense?
Almost always the chat template. A model prompted in another family's format will continue your text rather than answer it. Check that the template the app detected matches what the model expects.
Is my model uploaded anywhere?
No. Importing points the app at files already on your disk. Nothing is sent off the machine.
Try it on your own Mac
One signed, notarized download. No account, no token bill, and it keeps working with the Wi-Fi off.
Get Outlier for Mac