Support

Contact

hello@metalmind.app — anything about the application: bugs, questions, licence problems, feature requests. Include your macOS version, your Mac, and the model you were running. Better still, attach a diagnostic report.

For anything that is not about MetalMind itself — contracts, invoices, work enquiries — write to me@martinturek.dev instead. Both reach the same person; keeping them apart just means support does not get buried.

Sending a diagnostic report

In MetalMind, open Settings → System → Support. The report is assembled at that moment and shown to you in full before anything is sent. Read it, then press Email to open a pre-filled message. Nothing leaves your Mac until you send that mail.

The report holds the application and macOS versions, your chip, GPU core count and memory, the selected model and its state, your plan, server state, requests served, and the last error. It contains no conversation content.

Common problems

The model does not fit in memory

The model browser estimates the memory a model needs before you download it. If loading fails or the app evicts the model immediately, pick a smaller parameter count or a lower quantisation — a 4-bit 8B model needs roughly 5 GB of free unified memory, and macOS needs its own share on top of that. Quit other memory-heavy applications and try again.

The port is already in use

Another process is listening on 8080, the default. Either stop it, or pick a different port in Settings → Server — any free port works. Run `lsof -nP -iTCP:8080 -sTCP:LISTEN` in Terminal to see what holds it, and remember to update the base URL in your editor after changing it.

The editor will not connect

Check four things: the server is running in MetalMind, a model is loaded, the base URL in the editor ends with /v1, and the port matches. The API key can be any non-empty string. If the editor still fails, open the request inspector — if no request appears there, the editor never reached MetalMind and the problem is in the client configuration.

Apple Intelligence is not available

The system model only appears when Apple Intelligence is enabled in System Settings and has finished downloading its assets, on a supported Mac in a supported region and system language. If it is off, MetalMind runs downloaded MLX models exactly the same way without it.

There is not enough memory to train

Training needs considerably more memory than chatting with the same model, which is why the model step lists only what this Mac can actually train and marks the tight ones. If training stops for memory, go back to Model and pick a smaller one — your examples are kept. Quitting browsers with many tabs and other large applications first also buys you room.

Training is slow, or the Mac is on battery

Training runs the chip at full power. On battery macOS throttles it and the battery drains fast, so plug in before you start. The Mac getting warm and the fans spinning up is expected. You can close the wizard window — the run carries on and the top bar shows its progress.

A model cannot be trained

The model list marks why: CAN’T BE TRAINED means the architecture has no LoRA support in MLX, CHAT ONLY means the model works for conversation but not for training, ALREADY TRAINED means it is one of your own, and APPLE ONLY is the Apple Intelligence system model, which Apple does not expose for fine-tuning. Pick one marked TRAINABLE, or start from the recommended list.