This guide fills .qvac/models with both models Redact needs, so later runs work without a network. Do it once per checkout, online.

​
Download the OCR model

From the redact directory:
This reads public/test-files/redact/chat-private.png and prints JSON with blocks and stats. The OCR files (about 98 MB) land in .qvac/models.

​
Download the reasoning model

The decision model, Laya (laya-Q8_0.gguf, 450 MB), loads on the first full analysis. Run one:
  1. Start the service: npm run qvac:service.
  2. Start the app: npm run dev -- --port 3101.
  3. Analyse any test file in the browser and wait for Suggested redactions.
Alternatively, with no app running:
This loads both models and runs the four labelled cases.

​
Check the cache

You should see three files: one containing craft, one containing latin, and one containing laya. The offline commands refuse to start until all three are present.

​
Next