Settings and appearance
Settings has five sections: Model, Context length, Behavior, Connections and Appearance. The window scrolls on smaller screens.
Model
- Choose an installed local model or a model from a connected provider.
- Download models opens the curated catalog.
- Manage models lists local models, downloads a GGUF from a pasted HTTPS link, and removes unloaded models you no longer need.
- Open folder opens the models folder in File Explorer.
- Change and Reset switch to your own models folder or back to the default.
Context length
Choose 4k, 8k, 16k, 32k, 64k, 128k or 256k tokens.
- 8-bit KV cache (recommended) roughly halves KV-cache memory compared with a 16-bit cache; model weights and other allocations are separate.
- Multi-token prediction (auto) speeds up models with built-in MTP heads and does nothing on others.
Changing any of these reloads the local model. See Models & performance.
Behavior
- Start new chats with agent mode enabled sets the robot button's starting state for new chats.
- Show context meter in the top bar shows how full the context window is.
The Agent Mode default applies to new chats. Context-meter visibility updates across open windows, and you can click the meter for detailed statistics.
Connections
Add and edit AI providers such as OpenAI, OpenRouter or a custom server, and manage the service keys your Skills use. See AI providers.
Appearance
Pick a theme and a chat font size from 10 to 24 pt. The themes are Dark, Light, System (follows Windows), Ayu Mirage, Cobalt2, Dracula, Nord, Nord Light, One Dark, Rosé Pine Moon, Rosé Pine Dawn and Tokyo Night. The theme applies to every window and dialog. You can preview them on the home page.
Updates
Click ⓘ in the top bar, then Check for Updates. Checks are manual. If a newer release supplies an installer URL and SHA-256, Download and Install downloads it, verifies the saved file, saves conversations, closes LlamaBoss and starts the installer. The app reopens after installation.
Finish or stop active replies and tools before installing. You can cancel the download. If an installer and hash are unavailable, Open Download Page opens the trusted release link instead.
This updates the packaged application. There is no separate one-click updater for the bundled llama.cpp runtime in this source snapshot. A model newer than that runtime may require a newer LlamaBoss package.