Model profiles

Choose a model profile to show the language, recognition, cleanup, or voice controls that match the model your server runs.

Choice	What it tells Freehand	Example
Connection	Where to send requests and which backend API to use	A vLLM server and its authentication
Model	Which served model ID to request	Qwen/Qwen3-ASR-1.7B
Model profile	Which model behavior and controls to use	Qwen3-ASR

Names do not select profiles

For a manual connection, choose the profile yourself, including when the server uses a familiar model name or custom alias. Model discovery reads metadata; it does not download, run, or inspect a model to determine its behavior.

Choose a model profile

Manual connection

1. Open the workflow’s Settings cog. Choose Transcription in Voice or Audio file, Cleanup in either transcription workflow, or Speech in Text to speech.
2. Select the connection, then choose or enter its exact Model ID.
3. Review Model profile beneath the picker. Choose a specialized profile only when it matches the model your server runs, then choose Save.

Use Generic for standard backend options, including Whisper transcription, Kokoro speech, or your own cleanup instruction. When a backend offers only Generic, there is no profile selector. Manual whisper.cpp uses the model already loaded on the server.

Managed local connection

Choose a downloaded model in Local runtimes, then select its built-in Connection in the workflow. Freehand supplies the model ID and matching profile. Voice and audio files share the runtime’s transcription model. NeMo can also load a separate MagpieTTS model for speech.

See Local runtimes for available models and supported computers.

Workflow	Profile choice
Voice / Audio file	Generic, Nemotron, Parakeet, Qwen3-ASR, Cohere Transcribe, or Voxtral Mini Realtime
Cleanup	Generic for your instruction; S1-mini for its fixed prompt and trained controls
Text to speech	Generic, Qwen3-TTS on vLLM-Omni, or MagpieTTS on NeMo

With Nemotron, Qwen3-ASR, or Voxtral Mini Realtime, Voice’s Transcription options also offer Realtime transcription and optional overlay captions. Turning realtime off keeps the same connection and model for completed recordings. Audio-file transcription has its own selection.

Dedicated model guides

Nemotron 3.5 ASR streamingCompleted and live transcription, languages, and vocabulary boosting.

Parakeet TDT v3Completed transcription with automatic language detection.

Qwen3-ASRCompleted audio with context hints, plus optional live dictation.

Cohere TranscribeCompleted recordings and streamed file results in a selected language.

Voxtral Mini RealtimeLive or completed transcription with automatic language detection.

S1-mini by SuperwhisperEnglish cleanup with styling, structure, and context controls.

MagpieTTSLocal or remote speech with preset voices and language selection.

Qwen3-TTSSpeech with preset voices, languages, and style instructions.

Explore the visual model directory or compare speech model families before choosing a server.

Remember settings for each model

Freehand remembers options separately for each connection, feature, and model ID. You can choose a saved model without refreshing the server’s model list. A new model ID starts with Generic and default model options.

Remembered with each model	Kept when you switch models or connections
Transcription profile, context hint, temperature override, supported NeMo controls	Transcription language
Cleanup profile, output limit, reasoning override	Custom cleanup instruction and S1-mini style, structure, and context
Speech profile, voice, language, voice-style instructions	Speaking speed

Vocabulary, its Voice/file opt-ins, and Nemotron vocabulary strength are separate settings. Enable switches, recording behavior, shortcuts, and timeouts also remain feature or application settings. Running jobs keep the settings captured when they started. Model preferences contain no credentials or generated transcripts.

Read the model picker

Label	Meaning
Saved	Has remembered options
Server	Appears in current server discovery
Edited	Has unsaved options in this editing session
Manual ID	Comes from neither the saved list nor current discovery

An ID can have more than one label. Search or enter an exact ID, select it, then review its profile and options.

Save, switch, or discard

Action	Result
Switch models in the options sidebar	Changes the draft; returning to a model restores its unsaved edits
Save	Applies the current selection and all edited model options together
Save fails	Keeps applied settings active and preserves the draft for correction
Discard	Restores the applied selection and clears all model drafts
Switch connections	Restores that connection’s last selected model for this feature; a new connection needs a model chosen
Use first-run setup controls	Applies valid changes immediately, including restored model options, and shows save status

Review language after switching

Changing models preserves transcription language. S1-mini skips cleanup for known non-English input. For whisper.cpp, preferences belong to one Server-loaded model slot: review them yourself whenever you replace the model on the server.

Forget models or change a connection

Use the model actions menu’s Forget saved settings after saving or discarding edits. It removes the saved preferences and clears the current model selection. Enter the ID again to start from defaults. The limit is 32 models per connection and feature; forget an unused model if you reach it.

Connection change	Effect on model preferences
Rename or change authentication	Keeps remembered models
Change URL or backend profile	Clears remembered models and active model choices; select and review them again
Duplicate	Starts a separate set of preferences
Delete or remove an inactive use	Removes the corresponding remembered preferences