Local AI models

Amical can run entirely on your machine. No data leaves your computer, no account required.

Open Settings → AI Models to find three tabs: Speech, Language, Embedding.

Speech (Whisper)

The Speech tab lists every Whisper variant Amical supports, plus the Amical Cloud option. Each row shows speed (yellow zaps, 1–5) and accuracy (green dots, 1–5) ratings, plus the download size. Click Download on the model you want, watch the progress bar, then pick it from the Default Speech Model dropdown above the table.

Variants ship today: Whisper Tiny (~78 MB) → Whisper Base (~148 MB) → Whisper Small (~488 MB) → Whisper Medium (~1.5 GB) → Whisper Large v3 (~3.1 GB) → Whisper Large v3 Turbo (~1.5 GB). Pick what fits your hardware.

Already-downloaded models show a delete button instead of a download button. Deleting frees disk; you can re-download later.

Language (formatting & rewriting)

The Language tab pulls from local providers. Two are supported today:

• Ollama — install Ollama on your machine, pull a model (ollama pull llama3.2, etc.), and Amical syncs the list.
• OpenAI-compatible — point Amical at any OpenAI-compatible local server (LM Studio, vLLM, llama.cpp servers).

Pick a default from the Default Language Model dropdown at the top. The selected model is what the Formatting Alpha toggle in Settings → Dictation will use.

Embedding

Used internally for searching your transcripts and notes by meaning. Currently Ollama is the only local provider that exposes embedding models. Pick a default; you don't need to interact with this further.

Notes

• Whisper models on macOS run on Apple Silicon's Metal backend. Intel Macs work, just slower.
• The Preload Whisper Model toggle in Settings → Advanced keeps the model warm in memory for fast first-press latency.