Models and local compute

Manage local models

Open Model > Local > Transcription to see six Whisper models, their download state and the active selection.

1. Open the model list and request Download on a model that is not installed.
2. Wait for the inline download progress to finish.
3. Select an installed model to activate it. WhisperARC applies the change in the background.
4. Use the delete action to remove an inactive downloaded model when you need storage space.

Translation & Assistant models

Open Model > Local > Assistant to manage the Qwen model shared by local translation and dictation modes. This selection is separate from the Whisper transcription model.

Download a model, wait for it to be ready then select it. Qwen 3.5 2B is the installer default; 0.8B, 4B and 9B are also available. The displayed sizes are model downloads, so allow extra space for the runtime and working files.

Choose CPU or GPU

Choose CPU for the broadest compatibility or GPU when WhisperARC detects compatible hardware. The CPU thread slider stays available in GPU mode because those settings are used if GPU processing cannot start.

• CPU defaults to approximately half of the available logical cores on a new configuration.
• Increase CPU threads cautiously if transcription is too slow and the computer remains responsive.
• GPU can be selected only when WhisperARC detects a compatible graphics device.
• If the selected GPU cannot start, WhisperARC can use the configured CPU fallback.

When switching is blocked

WhisperARC can block a model or compute change while a dictation is awaiting a result or a media job is pending. This protects active work and queued progress. Finish or cancel the work then try the change again.