Live demo

Run a micromodel on your own machine.

A small specialist model downloads once, then runs entirely in this tab. Nothing you type is sent anywhere.

Model

Text in, the same text with PII masked out

What it does: Masks names, emails, phones, addresses, account and ID numbers, SSNs, birth dates and passwords, and leaves the rest of the text exactly as written. It runs in plain JavaScript on your CPU, so there is no GPU or big download.
Training: 530k text windows from Nemotron-PII, Gretel and generated notes. The whole model is a 4 MB file: hashed word embeddings plus 66 KB of 0.63-bit layers.
Accuracy: Leaves 0.2–0.5% of PII characters unmasked on held-out documents, in about 2 ms per note.

About this model

What it does: Masks names, emails, phones, addresses, account and ID numbers, SSNs, birth dates and passwords, and leaves the rest of the text exactly as written. It runs in plain JavaScript on your CPU, so there is no GPU or big download.
Training: 530k text windows from Nemotron-PII, Gretel and generated notes. The whole model is a 4 MB file: hashed word embeddings plus 66 KB of 0.63-bit layers.
Accuracy: Leaves 0.2–0.5% of PII characters unmasked on held-out documents, in about 2 ms per note.

Try an example

Your text

Redacts as you type. The model is 4 MB and runs on your CPU, so this one works on phones too.

Output

Loading the 4 MB model…

What you just saw

Private, accurate and free to run.

The network was used once, to download the model. Every run after that happened on your machine; turn off Wi-Fi and it still works.

The contact-notes model gets 137 of 142 held-out notes exactly right, and the jazz picks are level with frontier models. Both are 3B.

No per-call bill: the only cost was the electricity your computer used. That holds at a million runs a month.

We'll tell you if a micromodel can do it. Or book a demo, or email hello@micromodel.company.