The model does not arrive with the first page view
A responsible local AI experience should not silently start a large download. The Run Any AI tone reader waits until you choose Download & run locally. Only then does the browser request the model files.
Most of the download is model weights
Model weights are numerical values learned during training. They are not a database of prompts and they are not executable instructions by themselves. The current tone reader uses quantized weights, which store those values with less precision to reduce the download and memory footprint.
Its main weight file is roughly 67 MB. Tokenizer configuration and supporting runtime files bring the first-use transfer to approximately 70 MB, although browser and package versions can change the exact total.
The browser keeps a reusable copy
After download, the files are stored in browser-managed cache storage for this site. Later visits can reuse them, which is why the second run is normally much faster. Clearing site data, using private browsing, or pressing the lab’s removal control means the files may need to download again.
Local inference and network privacy are separate questions
The text entered into the tone reader is processed in browser memory and is not included in the model request. However, downloading the public files still creates a normal network request to the model host. That host can receive ordinary request information such as an IP address and browser details.
Local inference means the model processes your input on your device. It does not mean the page made zero network requests.
Try it with the labels visible
The live lab shows the estimated first download, progress, runtime, and cache-removal control before asking you to trust the claim.