Models
The model that writes your suggestions, and the AI models that rewrite dictation.
Tabussi uses a few kinds of model. All of the suggestion models run on your Mac.
| For | Where to pick it | Runs |
|---|---|---|
| Suggestions | Settings → Completions → Model | On your Mac |
| Dictation | Settings → Dictation → Models | On your Mac, or a server you choose |
| Rewriting dictation | Settings → Dictation → Models | On your Mac (S1-mini, Ollama, LM Studio) or a server you choose |
| Grammar (Labs) | Settings → Labs | On your Mac |
Models are stored in ~/Library/Application Support/Tabussi/Models.
The suggestion model
Settings → Completions → Model has a size slider. Tabussi marks the Best fit for your Mac's memory, and warns Heavy for N GB when a model would take more than half of it.
| Size | Model | Download | Best fit for |
|---|---|---|---|
| Fastest | Gemma 3 1B | 0.8 GB | Up to 23 GB of memory |
| Balanced | Gemma 4 E2B | 3.4 GB | 24 GB and up |
| Largest | Gemma 4 E4B | 5.3 GB | 48 GB and up |
Other Models has five more: Qwen3.5 0.8B, Qwen3.5 2B, Qwen 3 1.7B, Gemma 3 4B and Qwen3.5 4B. Downloads can be paused and resumed; delete one you no longer use from the Downloaded list.
Time from a key press to a suggestion, measured in the engine on a MacBook Air M3 with 16 GB:
| Model | Download | Typical | Slowest 5% |
|---|---|---|---|
| Gemma 3 1B | 0.8 GB | 13 ms | 15 ms |
| Qwen3.5 0.8B | 0.8 GB | 18 ms | 31 ms |
| Qwen3.5 2B | 1.3 GB | 23 ms | 38 ms |
| Gemma 4 E2B | 3.4 GB | 23 ms | 33 ms |
These are the engine's numbers; the app adds the time to read the field and draw the suggestion. Settings → Stats shows what you actually get.
A bigger model isn't always better here: on everyday English the larger ones didn't save more typing. Pick the recommended one unless you have a reason not to.
Gemma models are under Google's Gemma terms of use.
Memory and battery
- Unload When Idle frees the model's memory after 5 min, 15 min, 30 min (the default), 1 hour, or never. It loads again on your next key press. On battery the wait is at most 5 minutes.
- Save Power on Battery (off): on battery or in Low Power Mode, Tabussi suggests only when you press Suggest Now (⌃⌥⌘S), uses the smallest model you've downloaded in place of yours, and unloads sooner. Nothing is downloaded on battery.
AI models for rewrites
A dictation mode can have an AI model rewrite what you said. Keep any number of them; one is the default and a mode can pick another.
- OpenRouter. One key for hundreds of models, searched by name. Connect signs in through your browser and fetches the key; you can also paste one.
- Custom server. Any server that speaks the OpenAI API (
/chat/completions), by address. A server on your Mac, such as Ollama or LM Studio, keeps your words on your Mac. - S1-mini by Superwhisper. A small cleanup model (484 MB) that Tabussi downloads and runs itself. It removes filler and false starts, fixes punctuation and capitals, and writes numbers, dates, money and email addresses properly. Once downloaded it's the default rewrite. It takes no instructions; a mode using it gets a style instead. English only.
Add one with Add Model → OpenRouter Model or Custom Server. Each has its own Thinking (Off, Minimal, Low, Medium or High) and Temperature, both left to the model until you set them. Test sends a request and says how fast it answered. Keys are kept in your Mac's keychain.