A .kcppt template holds launcher settings plus download links for the models. When you launch it, KoboldCpp downloads the files it needs and starts with those settings. Templates are the quickest way to a working setup, including for images, voice and music.
A template is a config file meant for sharing. Most settings that depend on your PC are reset, so it works on other computers:
- The backend is chosen automatically for your hardware.
- GPU layers, threads, main GPU and tensor split are set to automatic.
- Passwords, SSL and Horde credentials are removed. Use MMAP, Use mlock and Direct I/O are turned off.
- Local file paths, Download Dir: and Device Override stay as they are. Check them before you share the template.
Load a template from the launcher
Section titled “Load a template from the launcher”- Click Get Help.
- Under "Or, Pick an Easy Template for Newbies", choose a category: Newbie Templates or Popular Templates.
- Choose a template in the second dropdown.
- Click Load Template. The launcher fills in the settings.
- Click Launch. The download starts now and can take a while.
Newbie templates
Section titled “Newbie templates”Sorted by how much graphics card memory (VRAM) they need:
| Prefix | Recommended VRAM |
|---|---|
| LowSpec | 6 GB |
| MidSpec | 12 GB |
| HighSpec | 24 GB |
| Template | For |
|---|---|
LowSpec-Chatbot, MidSpec-Chatbot | Chat |
LowSpec-ImageGen, MidSpec-ImageGen | Image generation |
LowSpec-MusicGen | Music generation |
HighSpec-CodingAgent | The KoboldCpp Agent for coding |
HighSpec-VideoGen | Video generation |
LowSpec-Chatbot, for example, loads Gemma 3 4B with vision, speech-to-text, text-to-speech and embeddings models, with a context size of 12288. Because that is below 16384, set KoboldAI Lite's Context Size to 12288 yourself; see Context size.
The templates live on Hugging Face: newbie templates and popular templates. The popular templates cover more models, such as Qwen3-VL 8B, Gemma 3 12B, and image and video models. More templates are in koboldcpp/kcppt.
Other ways to load a template
Section titled “Other ways to load a template”A .kcppt file loads like a .kcpps config:
- Load Config in the launcher.
- Drag the file onto the KoboldCpp exe.
--configwith a file name or anhttps://URL (koboldcppstands for your KoboldCpp file; see Command line):
koboldcpp --config https://huggingface.co/koboldcpp/newbie-templates/resolve/main/LowSpec-Chatbot.kcpptOn the command line, the backend is chosen automatically unless you add --usecuda or --usevulkan.
Where downloads go
Section titled “Where downloads go”KoboldCpp downloads every model field that holds a web address when you launch, not when you load the template. The one exception is Text Lora:, which must be a local file.
- Files are saved in Download Dir: on the Loaded Files tab (
--downloaddir). If that is empty, they go into the current working folder. When you double-click the exe, that is the folder the exe is in. - A file that is already there is reused, so the next launch starts without downloading.
- For text models split into several parts (
-00001-of-00003.ggufand so on), all parts are downloaded. - KoboldCpp downloads with
aria2c(bundled on Windows),curlorwget. If none of them works, it asks you to install one.
Create a template
Section titled “Create a template”Set everything up in the launcher, then go to the Extra tab and click Generate LaunchTemplate. The file is saved as .kcppt. The chat adapter and preload story files are copied into the template, so it works without them.
For a template other people can use, enter web addresses instead of local file paths in the model fields, for example https://huggingface.co/<user>/<repo>/resolve/main/<file>.gguf. HF Search fills in addresses in this form.
From the command line, --exporttemplate does the same, but does not copy the chat adapter and preload story files:
koboldcpp --model https://huggingface.co/<user>/<repo>/resolve/main/<file>.gguf --contextsize 16384 --exporttemplate mytemplateTo share a template or report a broken one, use the Discussions page of the popular templates repository.