Local models
TalkBar can do all of its work on your computer: transcription, speaker detection, translation, document search and, with Built-in AI, summaries and chat. Each of these needs a model on disk. Settings → Private Local AI is the one place to see, download and remove them.
Where to manage local models
Open Settings → Private Local AI. It has the Private AI Mode switch at the top and three tabs:
| Tab | What it holds |
|---|---|
| Local AI models | Small language models for Built-in AI: summaries, chat, polish, titles |
| Speech models | Speech-to-text models for live transcription, enhancement and dictation |
| System models | Helper models TalkBar downloads automatically when a feature needs them |
No models come with the installer. Each one downloads once, when you or a feature needs it, then works offline. Downloads come from Hugging Face and are checked before TalkBar uses them.
Local AI models
These models power Built-in AI. Each row shows the download size, speed and accuracy ratings, which features use it, and Download, Cancel or Delete.
| Model | Download | Memory | Good to know |
|---|---|---|---|
| Qwen3 0.6B | 380 MB | Small | Tiny and fast. The default for dictation polish. |
| Qwen3.5 2B | 1.2 GB | About 2 GB | Compact model for summaries on smaller computers |
| Qwen3.5 4B | 2.6 GB | About 3.5 GB | The default for summaries and other features |
| Qwen3.5 9B | 5.3 GB | About 7 GB | Stronger quality than 4B when you have the graphics memory; slower |
| Gemma 4 E4B | 5 GB | About 6 GB | Highest-quality Gemma 4 |
| Ministral 3 3B | 2 GB | About 4 GB | Compact multilingual model with strong structured output |
| LFM2.5 1.2B Thinking | 697 MB | Small | Fast, text only: for polish, titles and timelines, not for summaries or chat |
| LFM2.5 2.6B | 1.6 GB | Suited to computers with an 8 GB graphics card |
Larger models usually write better but need more memory and run slower. Speed depends a lot on your computer, so the best way to choose is to test.
Links at the top of the tab take you to Manage models (the Models tab in AI Providers) and Assign to AI features (the AI Features tab), where you choose which model each feature uses. See Providers and models.
Find my best model
Find my best model opens the Model Advisor. It runs a short test of each downloaded model on this computer, measures real speed and quality, and recommends the best default. It takes about a minute per model.
- The advisor shows your computer, for example Your system: 16 GB RAM · 8 cores.
- Download to test fetches a model you do not have yet.
- Best for your system names the recommended model. Click Make default to use it for AI features.
- If your hardware changed since the last test, the advisor asks you to run it again.
Local AI acceleration
| Option | What it does |
|---|---|
| Auto (GPU when available) | Uses your graphics card when one is available. Usually faster. Default. |
| CPU only | Runs on the processor only. Choose it if the graphics card causes problems or you want to keep its memory free. |
If Ollama is running and holding graphics memory, the built-in model can slow down. While Private AI Mode is on, TalkBar can warn you and offer to unload Ollama's model.
Speech models
The Speech models tab lists every speech-to-text model with its size and what it is used for: Live transcription, Enhancement or Dictation. Pick which model does which job in Settings → Voice (see Recording and transcription).
| Model | Download |
|---|---|
| Parakeet V3 (recommended for live transcription) | about 456 MB |
| Orukeet INT8 | about 641 MB |
| Whisper Small | about 465 MB |
| Whisper Turbo | about 1.5 GB |
| GigaAM v3 (Russian only) | about 215 MB |
| Canary 180M Flash | about 146 MB |
| Canary 1B v2 | about 692 MB |
System models
System models are small helpers. TalkBar downloads them automatically when a feature needs them; here you can re-download them or delete them to free space.
| Model | Download | What it does |
|---|---|---|
| Silero VAD | about 2 MB | Detects when someone is speaking before transcription |
| WeSpeaker speaker embeddings | about 25 MB | Recognizes who is speaking when transcripts are enhanced |
| Sortformer diarizer | about 470 MB | Separates speakers live during meetings |
| Sortformer diarizer (large meetings) | about 470 MB | Optional. Separates up to 8 speakers. |
| Nemotron 3 speaker detection (up to 8 speakers) | about 99 MB | Optional. Preferred by Automatic speaker detection for sessions with more than 4 speakers. |
| Translation models | about 400 MB per language pair | Live translation |
| BGE-M3 | about 417 MB | Semantic document search |
Hardware and disk space
TalkBar works without a graphics card. When one is available, Built-in AI can use it to run faster (see Local AI acceleration).
How much memory you need depends on the AI model you choose; the Memory column above is a guide. If you are unsure, start with Qwen3.5 4B, or a smaller model on a computer with little memory, and let Find my best model test what runs well.
Disk space adds up quickly:
- The Local AI models tab shows how much space your downloaded models use (… on disk).
- During first-run setup, TalkBar checks free space before it downloads the models you need for your first session.
- Models are stored in TalkBar's data folder, in its
voice/modelssubfolder. Open the data folder from Settings → Data & Backup.
Download and remove models
- Download fetches a model. You can keep working while it downloads, and Cancel stops it.
- Re-download replaces a damaged or incomplete copy.
- Delete removes a model and tells you how much space it frees. If a feature still uses the model, TalkBar warns you that the feature will need the model downloaded again to work.
You can download any deleted model again at any time.