# euRedact Cloud Is Open: the Public Beta Starts Today

> From today anyone can create a free euRedact Cloud account. The open-source engine keeps masking structured identifiers on your machine; the cloud adds the names, addresses and health data no pattern can describe. On a published benchmark of 900 documents in five languages it leaves 7% of personal data readable where the four other systems leave 33% to 48%. Here is what opens, what we measured, and how to start.

Markdown version of https://euredact.dev/blog/euredact-cloud-public-beta, generated from that page at build on 2026-11-19. Every page of the site is listed with its Markdown version at https://euredact.dev/llms.txt; the full SDK reference is https://euredact.dev/llms-full.txt.

[Back to blog](/blog)

 product-update cloud GDPR PII evaluation

# euRedact Cloud Is Open: the Public Beta Starts Today

*euRedact Team · 16 Nov 2026 · 9 min read*

**euRedact Cloud is in public beta from today: a free plan, no credit card. On a published benchmark of 900 documents in five languages it leaves 7% of personal data readable where the four other systems leave 33% to 48%, and masks 78% of the GDPR's Article 9 and 10 data where the others mask under 1% to 14%.**

Organisations across Europe sit on documents full of other people's personal data that they need to use, share and send to language models, and cannot risk leaking. Today euRedact Cloud opens as a public beta to solve that. Anyone can create a free account, no credit card, and the free plan comes with a launch offer that runs until 31 December 2026. The paid plans — Starter, Professional and Scale — go on sale on 14 December. Volumes, the offer and prices are on [the pricing page](/pricing), which is where they are kept current.

We have been building this since March, first as the open-source rules engine and then as the model behind it, and for the last two months with a closed group of alpha users who were generous with their time, their patience and their bug reports. We are genuinely excited to put it in front of a wider audience now. Not because the work is finished — a beta is a beta — but because the problem it solves matters to us. Handling personal data responsibly should not require a team that writes regular expressions for the identifier formats of thirty-one countries. That is the job we want to take off your hands.

## What opens today

Two things, which work together.

**euRedact Core** is the open-source SDK, for Python and Node.js, released under Apache 2.0. It runs on your own machine and finds the personal data that has a shape: national identifiers, IBANs, tax and VAT numbers, phone numbers, e-mail addresses and the rest of its 29 entity types, across 31 European countries, with the checksums that keep false positives rare. It was always free and it stays free; nothing about it is metered.

**euRedact Cloud** is the managed service that adds what no pattern can describe: a person's name, an address in running text, a job title, an employer, a diagnosis. It runs in the European Union, on Scaleway infrastructure in Paris; the machines that run the model have no public address. With the SDK in cloud mode, the identifiers the rules engine recognises are masked on your machine first, so the service only ever sees text with those already gone. By default nothing you send is kept: text and results are deleted as soon as the result is delivered, and only content-free usage statistics remain. A paid plan can set a retention window of up to 72 hours if you want one. The [trust page](/trust) describes exactly what happens to a document, the sub-processors are listed there, and the [Data Processing Agreement](/legal/dpa) is published.

Everything else that a contract needs is in place: the [General Terms](/legal/terms), the [privacy policy](/legal/privacy), and the [instructions for use](/docs/instructions-for-use) that say, per type of personal data, how well the service detects it and what your organisation still has to do with the output.

## What we measured

A redaction service should be judged on what it leaves behind, and we would rather be judged on documents that are not in our training data, with an answer key written under the benchmark's own guideline rather than our type definitions. So we are publishing our results on it alongside the launch: every type, set and language on [euredact.dev/benchmarks/cloud](/benchmarks/cloud), and the benchmark itself as [benchmark-2026-11-14 on GitHub](https://github.com/euredact/benchmark-2026-11-14).

One caution before the numbers. This is not an exact science. Whether a piece of text is personal data is often a judgement: does the name of a hospital ward, on its own, reveal a person's medical condition? Does a job title identify someone in a company of twelve? The benchmark's answer key takes positions on questions like these. Its guideline is not euRedact's own set of definitions, and we did not change the answer key to match ours. When our privacy philosophy differs from the benchmark, we accept the hit to our score: we deliberately mask context that lets a reader triangulate a person — departments, care settings and ordered tests in medical text — because a ward name beside a date can narrow a patient down as surely as a name. That is one of the reasons our raw precision is 0.85 rather than higher, and we would make the same choice again.

**900 test documents, in five languages.** Three synthetic sets of IT, HR and medical documents, 300 each, in English, French, German, Dutch and Polish, with 27,831 annotated personal-data items. None of these documents is in euRedact's training data.

**F2 0.921; the next best of the four other systems scores 0.706.** F2 counts recall twice, because a missed item leaks. euRedact 0.921, GLiNER2-PII 0.706, MS Presidio 0.656, bards.ai v3 0.632, Mistral Small 4 0.561.

**Leaves 7% of personal data readable; the others leave 33% to 48%.** euRedact leaves 2,041 of 27,831 items readable (7.3%). The four others leave between almost one item in three and almost one in two.

**Finds 84% of GDPR special-category and criminal data (Art. 9/10):** 87% in medical, 76% in HR and 68% in IT documents. The others: MS Presidio 42%, Mistral Small 4 11%, GLiNER2-PII 4% (not asked for these categories), bards.ai v3 2%.

**Consistent across languages:** F2 between 0.91 and 0.93 in all five (Dutch 0.911, English 0.919, French 0.924, Polish 0.924, German 0.929).

**Its widest lead is in medical text:** F2 0.900 on the medical set, where the others score 0.462 to 0.598. IT 0.931, HR 0.945.

**Rerun it yourself.** A 60-document sample with its gold standard, the harness, the configurations and our scores are on GitHub (source-available, for internal use). All 900 documents and their annotations are pinned by hash, and our full-set predictions are committed by hash, so they can't be changed after publication.

### The hardest categories

The special categories of Article 9 and the criminal-offence data of Article 10 — health, sex life, religion, political opinions, ethnic origin, offences — are the hardest part of this work. They rarely have a shape: a diagnosis is a phrase, not a format, and it can be an ordinary word in one sentence and a disclosure in the next. They are also the data that does the most harm when it leaks, and the data the GDPR holds to the strictest rules. That is where we have decided to put our effort, release after release, and it is where the benchmark shows the widest gap: 84% found against 42% at best for the four others. Measured the strict way, with an item counting only when at least three quarters of it is hidden, euRedact masks 78% of these items; the four others mask from under 1% to 14%.

Three of the four are general-purpose tools; the fourth, bards.ai v3, is built for European documents like we are, and deserves a word of respect: a small model that runs locally on a CPU and still reaches F2 0.632 gets a great deal out of very little. GLiNER2-PII was not asked for these categories at all. We built specifically for the European market — its identifier formats, its administrative languages and its strict privacy law — and for the one job that market needs done: reduce what a malicious actor can learn about a person from a document, while leaving the document as intact and usable as possible for the organisation that holds it. We would rather be the best answer there than an average answer everywhere. On the ordinary identifiers the gap is narrower. Leaving the Article 9 and 10 items aside, euRedact leaves 2.7% of the remaining 21,019 items readable, GLiNER2-PII 11.9%, bards.ai v3 26.0%, MS Presidio 31.3% and Mistral Small 4 33.6%. The special categories are where the approaches part ways.

> **When and how.** euRedact was measured on 14 November 2026 through the production service with the public SDK 1.0.0 (model release of 10 November). The other four were measured on 2 November 2026, with every configuration fixed before the sets were opened.
>
> **The data.** Synthetic documents, no customer data, written and annotated by language models under one guideline family. Agreement before adjudication was 0.85–0.97.
>
> **The method.** One run per system. "Finds" is recall with overlap matching: a span counts when any part of it was redacted. "Readable" means less than 75% of an item's letters and digits hidden.
>
> **Precision.** euRedact's precision is 0.85. Part of what it redacts beyond the annotations is masked on purpose, such as departments and tests in a medical context.

### Questions you may have

**Was euRedact tuned on these sets?** No document from them is in its training data. The release of 10 November did add newly written training examples for special-category data, after these sets showed where earlier releases missed it. The release frozen before that (1 November) scored nearly the same: F2 0.919, Art. 9/10 82%, 7.8% readable.

**How were the other systems set up?** Each was set up as a team would set it up, and frozen before any result was seen. MS Presidio with EU recognizers and term lists, configured in six rounds against 65 development documents from the benchmark's earlier sets, none of them among the 900 — a set-up standing in for about a week of a team's work. GLiNER2-PII with its model card's labels plus job title and money amount. Mistral Small 4 with a prompt that lists its 34 types, each with a definition, and asks for structured output. bards.ai v3 decoded as its model card does.

**What does 7% readable mean for my documents?** That detection is probabilistic and the output is not a guarantee: 2,041 of 27,831 items, 7.3%, still came through on this benchmark, more in some types than others. The [instructions for use](/docs/instructions-for-use) give the per-type figures and the checks we expect your organisation to keep in place.

## How to start

1. **Try it without an account.** The [live demo](/demo) runs on euRedact Cloud: paste a text, pick a country, and see what the rules engine and the model each find. It is for professionals evaluating the service, and runs on spare capacity.
2. **Create a free account** at [app.euredact.dev](https://app.euredact.dev/). When your account is ready we e-mail you a one-time code, which you enter on the page to sign in; the free plan needs no credit card and pauses at its limit until the next month.
3. **Install the SDK** — `pip install euredact` or `npm install euredact` — and follow the [quickstart](/docs/quickstart) (the source is [on GitHub](https://github.com/euRedact/euRedact)). Set your API key, add `mode="cloud"`, and the same call that masked identifiers locally now also returns names, addresses and health data.

## What a beta means

The service is live, monitored and under the General Terms, and [the status page](/status) shows its availability per day, measured by us every minute. What is still moving is the model: we will keep releasing it as the benchmark and your reports show where it misses. The instructions for use name the model release the published figures were measured on, and get a new version when a release narrows them.

**Found a document it handles badly? Write to [support@euredact.dev](mailto:support@euredact.dev)** — describe the shape of it, not the document itself.

Thank you to the alpha users who got us here. To everyone else: welcome, and tell us what you find.

## Try It on Your Documents

euRedact Cloud has a free plan, no credit card, and by default keeps nothing you send: text and results are erased on delivery. The live demo shows a redaction before you give an e-mail address.

[Try the live demo](/demo)[Plans and prices →](/pricing)

[All posts](/blog)
