The AI model
To make cards and chat, Lernbuddy needs an AI model. It is downloaded once (about 4.6 GB) and then runs entirely on the computer – no internet, no cloud.
Which model?
Lernbuddy uses Gemma 4 E4B by Google, a freely available language model, in a compressed version. It comes from Hugging Face, a well-known platform for AI models. The model is covered by Google's Gemma terms of use.
Downloading
On first start, Lernbuddy asks whether to download the model now. If you choose "Later", you'll find the download in the Help tab under AI model → Download model at any time.
- Wi-Fi only. On a mobile data connection (e.g. a phone hotspot that Windows treats as metered) the download is refused, so no data allowance is used up.
- Progress is shown in gigabytes and percent.
- Cancel and resume: A cancelled or interrupted download picks up where it left off on the next attempt.
Once "✓ Model installed and ready" appears, you're all set.
What works without the model
Even without the model, your child can practise existing cards: the quiz checks short answers without AI (see How answers are checked). Making cards, the chat and the feedback at the end of a quiz need the model.
What the download sends
The download is the only internet connection Lernbuddy ever makes. It fetches the model file and nothing else – no data about your child, no identifiers, no usage data. After that Lernbuddy works completely offline.
Performance
The AI runs on the computer itself. When the program starts, the model is loaded into memory; that takes a few seconds, so the first answer comes a little later. With a Vulkan-capable graphics card, answers appear smoothly word by word; on the processor alone it is noticeably slower.
Lernbuddy handles one AI request at a time. While a long chat answer is being written, for example, making new cards waits until it's done.