Private Local AI
Settings → Private Local AI is where you manage every AI model that runs on your computer and switch Private AI Mode on or off. It has three inner tabs: Local AI models, Speech models and System models. For what each model is good at, see Local models; for how the mode works, see Private AI Mode.
Private AI Mode
The top of the page shows the Private AI Mode switch.
| Setting | What it does | Default |
|---|---|---|
| Private AI Mode | Keeps all AI on this computer and turns off any external providers you connected. Features use the Private AI model assignments from Settings → AI Providers. You can't change it while a session is in progress. | Off |
The switch appears only after you add a cloud or command-line provider. Until then you see Private AI Mode is not available, because TalkBar already runs only on the built-in models and nothing needs restricting.
When you turn the mode on, TalkBar offers to download any local model it still needs. If a feature is still assigned to a model that isn't local, the switch shows Not fully local and a message tells you how many features to reassign.
If Ollama is holding GPU memory while the built-in model needs it, a warning Ollama is using the GPU appears with an Unload Ollama button.
Local AI models
The Local AI models tab (search entry: Local AI models) lists the built-in language models that run entirely on this device. These power summaries, chat, dictation polish, titles and knowledge extraction when they're assigned to those features.
| Setting | What it does | Default |
|---|---|---|
| Manage models | Opens the Models tab of Settings → AI Providers. | — |
| Assign to AI features | Opens the AI Features tab of Settings → AI Providers, where you pick which model each feature uses. | — |
| Find my best model | Opens Model Advisor. It tests your downloaded models on this computer (about a minute per model) and recommends the fastest one that passes a quality check. Set as default makes it your default model. | — |
| Local AI acceleration | Auto (GPU when available) uses your graphics card when it can. CPU only avoids using GPU memory. | Auto (GPU when available) |
Model list
Each model row shows its size, speed and accuracy ratings, and badges for the features that use it. Model Advisor results also appear here once you've run a test. The available models are Qwen3 0.6B, Qwen3.5 2B, Qwen3.5 4B, Qwen3.5 9B, Gemma 4 E4B, Ministral 3 3B, LFM2.5 1.2B Thinking and LFM2.5 2.6B.
| Control | What it does |
|---|---|
| Download | Downloads the model. The button shows the download size. While it downloads, the row shows progress and a Cancel button. |
| Retry | Appears when a download failed. Starts it again. |
| Re-download | Downloads a fresh copy of a model you already have. |
| Delete | Deletes the model file to free disk space. If a feature uses the model, TalkBar names the feature and warns that it will need a re-download to work again. |
The total space your downloaded models use is shown under the list.
Detect running Ollama
| Setting | What it does | Default |
|---|---|---|
| Detect running Ollama | While Private AI Mode is on, warns you when Ollama is holding GPU memory, so the built-in model can run at full speed. Unloading Ollama is always your choice. | On |
Speech models
The Speech models tab (search entry: Speech models) lists the speech recognition models used for live transcription, transcript enhancement and dictation. These models run on your computer.
Pick which model each feature uses in Settings → Voice. Here you download and delete them. Each row shows which feature uses the model: Live transcription, Enhancement or Dictation. The available models are Parakeet V3, Orukeet INT8, Whisper Small, Whisper Turbo, GigaAM v3, Canary 180M Flash and Canary 1B v2.
Each row has the same Download, Cancel, Re-download and Delete controls as local AI models.
System models
The System models tab (search entry: System models) lists helper models that TalkBar downloads automatically when a feature needs them. You can re-download them, download optional ones, or delete them to free space.
| Model | What it's for |
|---|---|
| Silero VAD | Detects when someone is speaking and splits the live microphone audio before transcription. |
| WeSpeaker speaker embeddings | Identifies who is speaking when a transcript is enhanced after a session. |
| Sortformer diarizer | Separates speakers live during recordings. |
| Sortformer diarizer (large meetings) | Optional. Separates up to 8 speakers live, for sessions with more than 4 speakers. |
| Nemotron 3 speaker detection (up to 8 speakers) | Optional. Automatic speaker detection prefers it when a session expects more than 4 voices; the large-meetings Sortformer model also handles up to 8. |
| Translation models (for example English → Russian) | Optional. Used for transcript translation, one model per language pair. See Live translation. |
| Semantic document search | Used to search your documents by meaning. See Documents. |
Related
Settings overview
How TalkBar Settings is organized, how to find any setting with search, and what highlighted, flagged and locked settings mean.
Projects, topics and keywords
Reference for Settings → Projects and Settings → Topics & Keywords, including project Trash, keywords cleanup and Action item labels.