InnoGE/deckname: the world's best AI model for German text anonymization

Why we trained InnoGE/deckname

This model was not born on a drawing board but in our projects. The pattern repeated itself for years: a team has a task a language model would finish in minutes. Sorting support inboxes, reviewing contracts, screening applications. Then comes the first look into the documents. Names, dates of birth, IBANs, customer numbers. By the time the data protection officer joins the conversation, the project is over.

And the objection is right. Handing such texts to a third-party cloud is exactly the risk the GDPR is meant to prevent. What happens next, we have seen often enough: the project is postponed, cut down until it helps nobody, or quietly run past data protection. We liked none of those outcomes. So we started training.

What surprised us most: nobody had trained this model before us. There is no shortage of anonymization tools, and yet the gap stayed open. A look at the market shows why.

What the market offers today

Text anonymization is not a new field. Anyone searching today finds five camps, and every one of them shares the same blind spot.

• The hyperscalers' cloud services. AWS, Microsoft and Google detect personal data as an API service, some with hundreds of entity types. But the text has to travel into their cloud for that, which is precisely where it must not go. And German plays a minor role: Amazon Comprehend, for instance, detects personal data in English and Spanish text only.

• Commercial specialist vendors. Some ship their detection as a container into your own data centre. You still pay by volume, and the models are generalists covering dozens of languages, not specialists for the fine grain of German text.

• Open-source frameworks such as Microsoft Presidio. Free and flexible, but a construction kit you have to assemble yourself. Detection depends on generic language components, and on German it drops off sharply: on German court decisions Presidio reaches 12.1 F1 in our measurement, InnoGE/deckname 94.5.

• Open PII specialist models. Small detection models such as the GLiNER family or Piiranha are considered the modern answer to Presidio. In our measurement they all fall short on German text: the best reaches 54.5 F1 on news text, and Piiranha, often recommended as the strongest multilingual model, misses over 90 percent of personal-data characters on German legal text.

• Prompted language models. The big frontier models come closest in quality, which is why we measured them most thoroughly. But they answer in seconds instead of milliseconds, cost tokens on every document, and above all they must see the raw data in plain text. Delegating your anonymization problem to a cloud AI does not solve it, it relocates it.

Five camps, one shared blind spot: a specialist model that truly masters German, covers English as well, runs fully on-premises and holds up on public benchmarks did not exist. For companies that take digital sovereignty seriously, that was the real gap. Today we are closing it.

How InnoGE/deckname works

InnoGE/deckname is not a chatbot and not another general-purpose model. We trained it for one sharply defined task: detecting personal data and replacing it with placeholders or pseudonyms. German and English run through the same model, trained on both languages jointly. In production it runs as a container inside your network, behind your firewall, on ordinary server hardware.

Two layers, two kinds of data

Personal data varies in how hard it is to find. An IBAN follows a fixed structure, a surname does not. That is why InnoGE/deckname treats the two separately:

• The rule layer finds structured items such as IBANs, tax IDs, card and phone numbers. Every hit is verified against its checksum. A string of digits that merely looks like an IBAN is not replaced, which keeps false alarms low.

• The trained model detects everything that needs context: personal names, addresses, organisations. Whether "Berger" is a surname, a district or a mountaineer, only the surrounding sentence can tell. That is exactly what the model is trained on, and nothing else.

Only together do they make a dependable system. Rules miss everything unstructured, and a model alone gives away the certainty a checksum provides on formatted data. Combined, the two layers cover what real documents actually contain.

What the measurements show

If you call your model the world's best, you have to prove it. So we sought out the comparison, against the strongest systems there are:

On public German benchmark corpora of news and legal text, InnoGE/deckname (v1) detects personal data more reliably than prompted frontier models, measured in July 2026 against GPT-5.6, Claude Opus 5 and Claude Fable 5. It answers in milliseconds instead of seconds, costs a fraction, and processes everything inside your own network.

Against Microsoft Presidio, the most widely used open-source tool, the gap is even larger: on the same corpora, Presidio lets 4.5 to 16 times more personal-data characters slip through. The open PII specialist models, among them Piiranha and two GLiNER variants, also trail far behind, even though we deliberately mapped their entity types onto ours generously. The best of them sits 32 F1 points behind InnoGE/deckname on news text and 73 on legal text. And where individual models shine with low leak rates, they buy it with over-redaction: on court decisions they mark court names and case numbers almost at random, and the text loses its value.

Then there is the fundamental point: a cloud model working as an anonymizer inevitably gets to see the unprotected raw data. InnoGE/deckname does not need that compromise, because it runs where the data already lives.

Why we also measure on real data

Public benchmarks are neatly curated and therefore a little too friendly. Real business text looks different: broken-off sentences, half addresses, names in places nobody expects, formatting that has grown over years. For training and evaluating InnoGE/deckname we therefore built our own dataset of real German business text, with a human-verified gold standard. Every model version has to prove itself against it, not just against the public corpora.

What companies use InnoGE/deckname for

InnoGE/deckname replaces none of your tools. It sits in front of them and clears paths that were blocked for good reason. Four of them come up particularly often.

As a gate in front of cloud language models

The most important use is also the simplest. InnoGE/deckname sits between your systems and the language model of your choice. Every text is cleaned before it reaches ChatGPT, Claude or any other model, and via the placeholder mapping you relate the answer back to the original data afterwards. Because the interface is OpenAI-compatible, a changed endpoint in your code is often all it takes. Quietly tolerated AI use becomes an official, documented process.

Customer service and inboxes

Support inboxes hold enormous automation potential, yet every single email contains names, addresses and contract numbers. With InnoGE/deckname in front, requests can be classified, summarised and routed automatically without customer data ever leaving the building. This works retroactively too: an existing inbox can be cleaned in one pass and analysed afterwards.

Legal and compliance

Contracts, files and briefs are among the most sensitive text a company holds. Cleaned, they can be searched, summarised and compared, including with tools that were previously out of the question. The audit report documents every single processing step along the way.

Data for analytics and your own training

Anyone who wants to use historical corpora for analyses, reports or training their own models needs them free of personal references. InnoGE/deckname makes such corpora usable without destroying their informational value. Structure and context survive; only the identities disappear.

Integration and operations: one container, one API, minutes to launch

A tool that sits in front of all your other systems must be neither a bottleneck nor a major project. That is why InnoGE/deckname ships as a single Docker container of around one gigabyte. One command starts it on the server hardware you already run; 2 to 4 CPU cores and 3 GB of RAM are enough, and a few minutes later the system answers its first request. Fully offline operation is supported as well.

The interface is OpenAI-compatible. Applications that talk to a cloud language model today talk to InnoGE/deckname tomorrow; in your code, little more than the address changes.

Your own servers are not a requirement, though. If you want to use the model without operating it, we serve it from our own AI infrastructure: on European servers run by a German provider, with no connection to US clouds, operated by the team that trained the model.

Existing corpora run through in one go instead of trickling through a paid API. And the step fits into pipelines where an extra second of latency per document would be unacceptable. How such an operation fits architecturally is covered in more depth in our piece on on-premises vs. cloud.

What InnoGE/deckname delivers, and what it does not promise

This is a field where the market likes big words. We deliberately avoid them. Terms like "anonymous" have a precise meaning in data protection law, and a promise a tool cannot keep will hurt you more in the critical moment than it ever helped before.

InnoGE/deckname detects and replaces personal data. In doing so it supports your anonymization and pseudonymisation processes and can be placed in your data protection concept as a technical and organisational measure under Art. 25 and Art. 32 GDPR.

Whether a result is anonymous or pseudonymous in the legal sense depends on context: on the residual information in the text, on the additional knowledge of the receiving party, on the purpose of the processing. That assessment rests with your data protection organisation, and no tool can take it off their hands. What InnoGE/deckname provides is the technical foundation: a traceable, repeatable, documented processing step that runs where your data lives.

Conclusion

The conflict between wanting to use modern AI tools and having to protect personal data does not resolve itself by waiting. It resolves at the point where texts leave the company. That is exactly the point we trained InnoGE/deckname for.

The result is a model that beats every system it faced on public German benchmarks, and is at the same time the first of its class to run entirely inside your own infrastructure. It works in milliseconds on ordinary CPUs, costs nothing per request, lets no data leave, and documents every replacement in an audit report. You can license it and run it yourself, use it in our AI infrastructure, or have us build it out into a finished solution.

The most convincing argument, though, is not our text but yours: paste an example, look at the result, judge for yourself.

FAQ: InnoGE/deckname and text anonymization