Custom local models

If you have a Whisper model that fits your work better than the standard ones — a
fine-tune on your domain's vocabulary, your accent, or a language the general models
handle poorly — you can import it and Zerm will treat it as a first-class local model.

What Zerm accepts

A .bin file in ggml format, the format whisper.cpp uses. That is the same
format as the models Zerm downloads for you.

If your model is in a different format — a PyTorch checkpoint, safetensors, a Hugging
Face repository — convert it to ggml first with the conversion scripts in the
whisper.cpp project. Zerm does not convert anything.

Quantised ggml models work and are usually the better choice: much smaller, with a
small accuracy cost.

Importing

1. Open Models, under Dictation in the sidebar, and choose the Local filter.
2. Scroll to the end of the list and select Import Local Model…
3. Choose your .bin file.

The model is copied into Zerm's model directory and appears at the end of the Local list,
marked "Imported local model". Select it like any other model, set it as default, use it
in Transcribe File, or pin it to a Power Mode.

If a model with the same file name is already imported, nothing is copied; rename the
file first.

Delete it from the card's menu with Delete Model. That removes Zerm's copy; your
original file is untouched.

What to expect

It runs locally. Same as any bundled Whisper model — no key, no network, Metal
acceleration on Apple Silicon.

Everything downstream is identical. Filler-word removal,
dictionary replacements and vocabulary, formatting, Output Format,
and enhancement all behave the same. The only thing that changes is
what produces the first draft.

Language support comes from the model. Zerm cannot know what your fine-tune covers,
so it offers the full Whisper language list and files the model under Multilingual.
Choose a language the model actually handles, or leave it on Auto-detect.

Size and memory are on you. The memory warnings and Recommended picks Zerm shows for
catalog models are based on sizes it knows in advance, so an imported model gets
neither. One of the same footprint has the same requirements; a very large one will be
slow or will not load.

If it does not work

• The file is rejected. It is not ggml. Convert it.
• It loads but produces nonsense. Usually a language mismatch, or a model trained
  for a task other than transcription.
• It is much slower than a bundled model of similar size. Check whether it is
  quantised. An unquantised model is several times the size and correspondingly slower.

Custom cloud models are separate

This page is about local files. To point Zerm at a hosted transcription API, use the
Custom filter on the Models screen and Add Model instead — that path accepts any
OpenAI-compatible transcription endpoint. See models.