docs/MODEL_STORAGE_POLICY.md

Model Storage Policy

Are the LLMs on the key?

Yes. Shipping product: the LLMs are already stored on the USB/SSD before the customer receives it.

Windows mode:

• Ollama model blobs live under models/ollama on the drive.
• Optional GGUF files live under models/gguf when a local llama.cpp server is supplied.

Linux boot mode:

• The same model directories are copied or mounted under /opt/JackAILocal/models/*.

Are models downloaded when the key is inserted?

No. Customer first run must not download models. The product must work offline and in airplane mode.

Downloads happen only in these cases:

1. Factory preparation before shipment.
2. Explicit signed update pack.
3. Explicit advanced model-management action while online.

Why?

Because the product promise is offline, no-install, no-admin, non-technical. A customer who inserts the drive should not hit a 4 GB to 20 GB download, a failed Hugging Face token, a firewall block, a slow hotel network, or a broken model pull.

Runtime flow

At first launch, the launcher sets:

OLLAMA_MODELS=<drive>\models\ollama

Then Ollama reads model blobs from the drive. Ollama officially supports changing the model storage directory with OLLAMA_MODELS; this is what makes the portable model directory possible.