Omit Redact 1.0 is stable on Windows

Omit Redact 1.0.0 is out for Windows 10 and 11. It is our first stable release, after seventeen betas, and it is the build we would now tell anyone to install, not just people who enjoy filing bug reports.

Stable is a promise about what we will not change under you. Version numbers from here follow the usual rules, the download page always serves the current stable build, and nothing went into this release at the last minute. 1.0.0 is the build our final beta grew into, with the beta label taken off.

If you have not followed the betas: Omit Redact finds and removes personal data from PDFs, scanned images, spreadsheets and plain text, entirely on your own machine. Your originals are never modified, because the redacted copy is written beside them. Redactions can be one way, or reversible through an encrypted local vault, so you can send a redacted file out and restore the names when it comes back.

The headline change is inside the Accuracy tier. It now runs the first detection model we have fine-tuned ourselves. Until now, Accuracy ran an open model as published. It was a good starting point, but it missed things we did not want missed and flagged things that were plainly not personal data, and both of those cost you time.

The second one matters more than it sounds. A model that flags everything is not cautious. It teaches you to click past its warnings, and the one real name it caught gets clicked past with the rest.

We measured the new model the way we would want a vendor to measure theirs: on test sets it never saw during training, with the same detection threshold the app uses, and against the model it replaced, in the same run. The number we care about most is the share of personal data that nothing catches at all, because a missed name is the mistake that leaves a document unsafe. A name caught under the wrong label still gets redacted. A name nobody caught does not.

On NVIDIA's Nemotron-PII test set, which is public and which we never trained on, the open model missed 22.5% of the personal data across 5,000 documents. The new model misses 0.8%. Those figures are for the model on its own; in the app, the pattern and checksum recognizers run alongside it.

False alarms fell just as far. We ran both models over 744 documents that contain no personal data at all. The open model flagged something in 567 of them, 76%. The new model flags something in 88, 12%. That is the difference you will notice first: far less to dismiss on every review screen.

You may wonder why we do not quote a score on ai4privacy, the PII dataset most vendors publish numbers on. Part of our training data comes from that family of datasets, so a score there would be grading our own homework. We have said before that in-distribution scores are not evidence, and that applies to us too. Nemotron-PII is the cleanest outside test we have, so that is the one we lead with.

None of this means the model catches everything. 0.8% is not zero, and on harder material, financial documents especially, it misses more. That is why the review step exists, and why the Fast tier still validates structured identifiers like card numbers and IBANs by their checksums instead of trusting any model to recognise them. Use both, and review anything that matters.

This is the first fine-tuned model we have shipped, not the last. The next version is already in training, aimed at fewer misses again and better accuracy overall, and it will arrive as an ordinary update when it beats this one on the same tests. If it does not beat this one, it does not ship. We already held one candidate back this month for exactly that reason.

The other thing people keep asking about is the Mac. The macOS build is in final testing, on both Apple silicon and Intel, and we expect to ship it this week. It will get its own post when it is out.

There are still two Windows installers, and the download page lists both with their measured sizes. The offline installer carries all seven Fast tier languages and the Accuracy tier, and never touches the network. The light installer is a much smaller download that starts in English and fetches the other languages when you ask for them. It does not include the Accuracy tier yet. Downloading Accuracy from inside the app is on the way, so if you want the new model today, take the offline installer.

One thing has not changed since the beta. We have not bought a code signing certificate yet, so Windows SmartScreen will warn you the first time you run the installer. The warning is about the missing certificate, not about anything in the file. If you would rather check than trust us, install it, disconnect from the network, and redact a file end to end. Nothing in Omit needs the internet to do that. The download page shows the exact screen and which button to press.

Protect, the clipboard guard, stays free for everyone. Everything paid, which is file redaction, re-identifying from the vault, and the Accuracy tier, comes with a 15 day trial that starts only when you ask for it, not when you first open the app.

Bug reports go to the issues tracker on the releases repository. Please include the version string from Settings. If anything in this post turns out to be wrong, tell us, and we will correct it in the open instead of quietly editing around it.