Models and hardware
Minutist runs its models on your machine. Each is a download you start on first use, the same network call the free build’s promise refers to; nothing else is sent.
What gets downloaded
Section titled “What gets downloaded”| Model | What it does |
|---|---|
| Speech-to-text (Parakeet) | Transcribes your meetings, with per-word timestamps and speaker labels. |
| Summariser (local LLM) | Produces summaries, answers questions in chat, and drives the live assistant. |
| Vision model | Reads text in image attachments (OCR). No extra download, included in the initial set. |
| Text-embedding model (BGE-M3, ~600 MB) | Indexes your attachments and transcript so chat and the live assistant can retrieve the relevant passage on demand rather than reading whole documents. Downloaded when you first use retrieval. |
Exact model names and file sizes are shown inside the app because they change between releases.
Hardware requirements
Section titled “Hardware requirements”Most of what Minutist does runs on CPU. A discrete GPU is faster but not required, with one exception.
Live assistant automatic mode
Section titled “Live assistant automatic mode”The live assistant reads the transcript as a meeting runs and keeps a running digest of it. Its Auto mode, which engages the assistant on its own, needs a discrete GPU: on a machine with an integrated GPU only, Auto resolves to off. The live assistant is off by default and opt-in.
Everything else (transcription, speaker labelling, summarisation, retrieval, OCR) runs on CPU. A discrete GPU speeds these up but is not required.
Model storage
Section titled “Model storage”Model files are stored in your local data directory (XDG_DATA_HOME/ai.minutist/ on Linux, %LOCALAPPDATA%\Minutist\ on Windows, ~/Library/Application Support/ai.minutist/ on macOS). They are not re-downloaded after the initial fetch unless you clear them.