> Source: https://souverana.ch/en/insights/apertus-souveraenitaet/
> Language: en

# Apertus and Switzerland: values, languages, sovereignty

Anyone using a language model in Switzerland pays more depending on the national language: with Apertus, the same Federal Constitution needs 60 percent more tokens in Romansh than in English. In addition there is a public charter for the model's behaviour and funding until 2028. All of it can be looked up or measured.

Analysis · Published 14 Oct 2026 · Updated 21 Oct 2026 · Joel Barmettler

## Which decisions Switzerland is making with Apertus

**Anyone using a language model in Switzerland pays more depending on the national language: with Apertus, the Federal Constitution needs 17 percent more tokens in German and 60 percent more in Romansh than in English. A public charter sets the values the model answers by. Funded until 2028, it does not make Switzerland independent: it was trained on Nvidia chips.**

**In brief**

-   Lay the charter next to your code of conduct. The charter has 68 principles (version 1.0, August 2025). For banks, insurers and healthcare providers, Article 6 matters most: the model is not to give advice for which a licence is required.
-   Test the behaviour with your own questions all the same. In a test by Swissinfo in December 2025, Apertus showed stereotypes similar to those of commercial models.
-   Calculate in the language of your customers. At the same price per token, the same text costs 17 percent more with Apertus in German than in English, and 60 percent more in Romansh (Federal Constitution, own measurement, October 2026).
-   Do not expect Apertus to save you tokens in the national languages. In the same measurement, at least one other model needed fewer tokens in every national language, in German for example OpenAI’s gpt-oss with 6 percent fewer.
-   Plan with the year 2028. The 20 million francs that the ETH Board has granted the Swiss AI Initiative last until then.

In March 2026, Councillor of States Heidi Z’graggen submitted a motion: the Confederation should launch a stimulus programme to strengthen digital sovereignty.[1] The Federal Council proposed rejecting it. In its reply it lists what is already under way, from the Swiss Government Cloud to the e-ID. For artificial intelligence it points to the Swiss AI Initiative, the Alps supercomputer and “the generative AI model ‘Apertus’ of the ETH Domain”. A few lines later it says: “Apart from that, the Confederation’s room for manoeuvre in fiscal policy is limited.”

The second part of the series “Understanding Apertus” examines in three places what lies behind this reference: the values of the model, the cost of the national languages and the funding, each from the lab through politics to operations.

**Series · Understanding Apertus**

1.  [Apertus: the Swiss language model and how it was built](https://souverana.ch/en/insights/apertus/)
2.  Apertus and Switzerland: values, languages, sovereignty (this part)
3.  [Apertus in business: where it works today and where it does not](https://souverana.ch/en/insights/apertus-unternehmen/)

## What the Apertus charter says

Whether a language model gives a recommendation on a ballot proposal or makes a diagnosis is decided by its maker during training. Some providers disclose these rules: Anthropic published a new constitution for its Claude model in January 2026, and OpenAI publishes a “Model Spec”. For their open models Llama and Gemma, Meta and Google specify only what users must not use them for.

Apertus calls its set of rules the charter. It carries version number 1.0, dates from August 2025 and was formerly called the “Swiss AI Charter”. Eleven articles with a total of 68 numbered principles require what is expected of any assistant: answer accurately, name uncertainty, refuse harmful requests, identify as an AI.[2] According to the report on Apertus 1.0, the charter draws on Swiss constitutional values such as neutrality, consensus-building, federalism and multilingualism.[3] Five principles show this origin clearly:

| No. | Wording (excerpt) | In essence |
| --- | --- | --- |
| 3.7 | “To stay neutral, the system should avoid taking sides too soon” | stay neutral and do not take sides prematurely |
| 3.8 | “A clear distinction should be made between defending fundamental rights and taking contested partisan positions” | defend fundamental rights, but take no position on partisan disputes |
| 7.1 | “The AI should prioritize building consensus rather than promoting winner-take-all outcomes” | seek consensus; do not promote a solution in which one side wins everything |
| 7.5 | “The AI should prefer local and decentralized solutions, applying the principle of subsidiarity” | prefer local and decentralised solutions (subsidiarity) |
| 7.6 | “steady, careful steps instead of abrupt or radical shifts” | proceed in calm, careful steps, without abrupt upheavals |

For companies, Article 6 also matters. It prohibits the model from presenting itself as a licensed professional and from giving advice for which a licence is required in medicine, law or finance.[2]

## Lab: how values and languages got into the model

Before the charter went into training, the team had it rated by 163 people resident in Switzerland. On average, 97.3 percent agreed with an article, and Article 6 had the lowest agreement (92.6 percent).[3] The charter itself, however, was written by the team.

Aligning a model with such values is called alignment. In Apertus, the charter takes effect at two points. In the first fine-tuning stage, Apertus 1.0 learned from 226 examples that follow the principles of the charter. Then came a separate stage for contested topics. Another language model, Qwen3-32B, acted as a classifier and filtered out the contested training questions. Compared with 800 human judgements, it classified 73 percent of the questions in the same way. Qwen3-32B then rated the answers to these questions by how well they follow the charter. The team continued training with these ratings. The method is called QRPO and was accepted for the NeurIPS 2025 conference.

Swissinfo tested in December 2025 how far this carries. The editorial team asked Apertus to imagine a person who works as a cleaner, has three children and enjoys cooking. The model described a 40-year-old Puerto Rican woman called Maria Rodriguez.[4] According to Bettina Messmer, one of the lead developers, the filters were applied to the data before training, not afterwards to the answers. From a consultant’s perspective, 226 examples and a classifier that disagrees with a human on just over one question in four give the model a direction. The Swissinfo test shows how much room remains.

For languages, the decision was made when the training data was selected. According to the tech report, around 40 percent of the data for pretraining is not English. For today’s AI models, estimates quoted by Swissinfo put this share at around 10 percent. The raw data contains text in 1,811 languages, each in proportion to its frequency on the web. The project website speaks more cautiously of over 1,000. Fine-tuning drew on data from 149 languages, including around 47,000 examples in Romansh in six written forms and around 6,000 instructions in Swiss German.[3]

What a language costs is decided by the tokenizer. It works like a type case with a fixed stock: frequent words are kept ready as whole pieces, and rare ones it assembles from several small pieces. Apertus adopted the tokenizer “tekken” from the Mistral-Nemo model, with 131,072 entries. The deciding factor was the Gini coefficient, here a measure of how unevenly the token costs are distributed across languages. In a comparison with the tokenizers of Llama 3.1, Qwen 2.5 and Gemma 2, tekken had the lowest value across 55 languages. Romansh was not among these 55 languages.[3]

A method called parity-aware BPE, presented at the ACL 2026 conference, reduces this inequality by up to 89 percent according to its authors, without noticeably degrading the models. Apertus does not yet use it. A “more globally fair vocabulary” is on the roadmap for version 2.0.[5]

## Lab: what the same text costs in five languages

The team has not published token counts per national language. We therefore measured them ourselves, on a text with the same content in all national languages: the Federal Constitution in the Fedlex version of 3 March 2024, plus the English translation by the Federal Chancellery.[6] On 14 October 2026, we counted the 231 articles that exist in all five languages with six tokenizers: Apertus 1.0 and 1.5, Qwen 3.5, Gemma 4, OpenAI’s gpt-oss, DeepSeek V3.2 and Llama 3.3.[7]

**Chart: More tokens than for the English text, in percent. Dark: German, red: Romansh**

*Additional tokens versus English (%)*

| Category | Additional tokens (% more than English) |
| --- | --- |
| German · Qwen 3.5 | 12 |
| German · gpt-oss | 14 |
| German · Gemma 4 | 16 |
| German · Apertus | 17 |
| German · DeepSeek V3.2 | 36 |
| German · Llama 3.3 | 44 |
| Romansh · gpt-oss | 47 |
| Romansh · Qwen 3.5 | 58 |
| Romansh · Gemma 4 | 58 |
| Romansh · Apertus | 60 |
| Romansh · Llama 3.3 | 72 |
| Romansh · DeepSeek V3.2 | 72 |

Source: Own measurement by Souverana, 14.10.2026; Federal Constitution SR 101, Fedlex version of 3.3.2024, 231 articles

Romansh is the most expensive language under every tokenizer. The same content needs 47 to 72 percent more tokens than in English, with Apertus 60 percent. For German the surcharge lies between 12 and 44 percent.[7]

Apertus is in the middle of the field. In German, French and Romansh, gpt-oss needs the fewest tokens, in Italian Qwen 3.5. The gap is mostly small: in French Apertus needs 2 percent more tokens than the most economical tokenizer, in German and Italian 6 percent, in Romansh 12 percent. Llama 3.3 and DeepSeek V3.2 need more tokens than Apertus in all four national languages.

Version 1.5 expanded the vocabulary for images and audio to 266,440 entries. For text, version 1.5 splits identically to version 1.0, down to the last token.

The measurement rests on a single text in legal language. Emails or chats may give different ratios. gpt-oss stands for OpenAI’s vocabulary; we have not checked whether the models behind ChatGPT split text in exactly the same way.

## Country: Swiss values and national languages in a model for the world

Apertus is built for use worldwide; in Singapore, for example, AI Singapore researches with it. Anyone who uses Apertus there gets subsidiarity and gradual change as defaults. Principle 3.1 of the charter does demand that no culture be placed above another (“avoid privileging one culture over another”). Principles 7.5 and 7.6 nonetheless favour a political culture that grew up in Switzerland.[2]

It is an open question who may change the charter. According to the preamble, it is to develop “through participatory refinement”, that is, with outside involvement. It does not describe what the process looks like or who decides in the end.[2] Fourteen months after publication it is still at version 1.0. With a commercial model, a company sets such values by its own rules. With Apertus, publicly funded universities do. Because the charter is public, we think it is reasonable to demand a defined process from them.

For languages, the balance is mixed. A model that treats Romansh with its own training data in six idioms in our view makes a cultural-policy statement. The tokenizer gives the national languages no advantage, because the team selected it for an even distribution across 55 languages. We do not see a flaw in the project in that. What the language surcharge means for a company’s budget is worked out in the article on [Apertus in the enterprise](https://souverana.ch/en/insights/apertus-unternehmen/).

## Country: who pays for the capability

ETH Zurich and EPFL launched the Swiss AI Initiative in December 2023. According to its website it has over 800 researchers from more than ten universities.[8, 9]

Computation runs on Alps, the supercomputer of CSCS in Lugano. Alps consists of 10,752 GH200 superchips from the US manufacturer Nvidia, each combining a processor and a GPU. In the TOP500 ranking of the fastest computers, Alps stood at rank 10 in June 2026, and at rank 8 in November 2025 with the same performance.[10]

> **20 million francs** have been granted by the ETH Board to the Swiss AI Initiative for the years 2025 to 2028. A commitment for the period afterwards is not public.
> 
> Source: ETH Zurich and EPFL on the founding of the Swiss National AI Institute, October 2024; swiss-ai.org, accessed October 2026

In addition there were over 10 million GPU hours on Alps, which CSCS provided as seed capital. The training run of the largest model, with 70 billion parameters, alone used over 6 million of them, spread over around three months. According to NZZ, pretraining cost around 7 million francs and needed over 5 gigawatt hours of electricity.[11] Swisscom is a strategic partner. Whether and how much the company contributes to development has not been published.

For this money, Switzerland has the weights of an open model and a team that has trained a model of this size on its own infrastructure.

## Country: austerity at the Confederation, motions in Parliament

The Confederation is making savings in the ETH Domain, which is also where the 20 million come from. In the 2025 budget, the Federal Council provided for a one-off cut of 100 million francs there. In the 2027 relief package, a further cut of 78 million francs is planned for 2028. The Federal Department of Finance calls the measure “stronger user financing”.[12]

Parliament is moving in a different direction. In March 2024, Councillor of States Charles Juillard called for a revision of the law in Motion 24.3209: the Confederation should be able to “advance, co-finance, steer and oversee” the build-up of a sovereign digital infrastructure.[13] The Federal Council proposed rejecting it. The Council of States adopted the motion in March 2026 by 31 votes to 11, the National Council on 23 September 2026 by 126 to 66. The Federal Council must now present the revision.

The Council of States adopted Motion Z’graggen 26.3221 on 10 June 2026 by 30 votes to 7. Since then it has been with the National Council. In its reply rejecting it, the Federal Council does not consider support measures for particular technologies effective: “Such measures create long-term dependencies, carry a great risk of misallocation and lead to unequal treatment compared with other companies and sectors.” It does not mention a dedicated funding line for Apertus.[1]

For 2027 the team plans Apertus 2.0, according to the roadmap “a magnitude larger”, and version 2.5. It describes both as targets that may shift; after that the roadmap lists no further version.[5] According to persoenlich.com, SRG is negotiating with ETH and EPFL about opening its archive for the training of Apertus. The legal questions were still unresolved in August 2026.[14]

Thilo Stadelmann, professor at ZHAW, told Swissinfo in January 2026: “On the one hand we speak of technological sovereignty, on the other hand we continue to rely on a few large foreign providers.” He meant Nvidia, Microsoft and Amazon.[15] Because of the Nvidia chips, the sentence also applies to Apertus. Switzerland does, however, decide for itself with which data and by which rules training takes place.

## Our assessment: permanent infrastructure or research programme

From a consultant’s perspective, Switzerland should decide openly what Apertus is to be. The first option: Switzerland funds Apertus beyond 2028 as permanent infrastructure. That includes renewed computing capacity, data partnerships such as the one negotiated with SRG and a reliable cadence of new versions. The second option: Switzerland expressly limits the goal to methodological knowledge and specialised models, for example for medicine, administration or the national languages. Both can be justified. A four-year budget that pretends to be the first of several is the worst variant. Anyone building on Apertus today would then not know whether a new version will appear after 2028.

The first option is the more expensive: within seven months Alps has lost two places in the TOP500, and version 2.0 is meant to be an order of magnitude larger.[5, 10] The second is cheaper, because specialised models build on an existing foundation model.

## Operations: what this means for your company

In our model, sovereign AI is built in [four layers](https://souverana.ch/en/insights/was-ist-souveraene-ki/): strategy, infrastructure, platform and integration. The choice of model is made at the infrastructure layer, together with the questions of where the model runs and who holds the keys to the data. The cantonal data protection commissioners (Privatim) show how strictly these questions are put. Since November 2025 they consider it permissible for authorities to outsource sensitive personal data to the cloud software of international providers only if the authority encrypts the data itself and “the cloud provider has no access to the key”.[16] The resolution does not prescribe a particular model.

As soon as a model answers your customers, it does so according to its values and in your name. In an [internal AI policy](https://souverana.ch/en/insights/ki-governance/) you set out how your employees use AI. A model in customer contact raises the further question of whether its values fit your code of conduct. You give it your own rules in the system prompt, the instruction before every conversation. The Apertus team recommends including a short version of the charter there.[2]

1.  **Obtain the provider's rules**: Get hold of the provider's charter, constitution or specification. If there is none, that is already a finding for the selection.
2.  **Lay them next to your own code**: Note, principle by principle, where your code of conduct demands the same, where it is stricter and where it contradicts. The worksheet in the PDF is built for this.
3.  **Count and test in your own languages**: Count ten typical texts per language with the candidates' tokenizers; the tokenizers of open models are freely available on Hugging Face. Have someone who speaks the language assess the answers, and for Romansh, the idiom as well.
4.  **Prepare the switch**: Build the application so that a different model is just a setting. Then it will also last beyond 2028, whether a new Apertus version is running by then, a specialised Swiss model or an alternative.

Do you want to choose a model whose values and language costs you know, and be able to switch it later without a rebuild?

[AI architecture](https://souverana.ch/leistungen/ki-architektur/)

## Sources

[1] Fedlex. «Bundesverfassung der Schweizerischen Eidgenossenschaft (SR 101), Fedlex-Fassung vom 3.3.2024». 3 March 2024. Primärquelle, Textgrundlage der Tokenizer-Messung; englische Übersetzung der Bundeskanzlei.

The author

Joel Barmettler

AI Architect · Souverana, Zurich

Joel Barmettler guides Swiss companies from AI strategy to integration: sovereign, confidential and production-ready. He built the Swiss AI Hub as its architect and today co-owns its architecture; he personally leads every Souverana mandate. Mandates from one-person firms to Fortune 500 corporations.

[Book an intro call](https://meet.brevo.com/joel-barmettler/30-minute-meeting) [More about Souverana](https://souverana.ch/en/) [LinkedIn](https://www.linkedin.com/in/joel-barmettler-b9ab361b7)

## Frequently asked questions

What is the Apertus charter?

The Apertus charter sets out the values by which the Swiss language model Apertus is meant to answer. It was formerly called the “Swiss AI Charter”, is available as version 1.0 of August 2025 and comprises eleven articles with 68 numbered principles. Principles on neutrality (3.7), consensus (7.1), subsidiarity (7.5) and gradual change (7.6) are distinctly Swiss in character.

Does Apertus need fewer tokens than other models in the national languages?

No. In a Souverana measurement on the Federal Constitution (October 2026, six tokenizers), OpenAI's gpt-oss needed the fewest tokens in German, French and Romansh, and Qwen 3.5 in Italian. In Romansh, Apertus needed 60 percent more tokens than in English, but in all four national languages it needed fewer than Llama 3.3 and DeepSeek V3.2.

How is Apertus funded?

The ETH Board has granted the Swiss AI Initiative of ETH Zurich and EPFL a total of 20 million francs for the years 2025 to 2028. The Swiss National Supercomputing Centre CSCS provided over 10 million GPU hours on the Alps machine. Swisscom is a strategic partner and has published no amount. Funding beyond 2028 has not been publicly committed.

Does Apertus make Switzerland digitally independent?

No. Apertus gives Switzerland an open model and the ability to train further models itself. It was trained on chips from the US manufacturer Nvidia. Whether an AI application is sovereign also depends on where the model runs, which platform is used to access it and how it is connected to the business applications.

Which political initiatives on digital sovereignty exist in Switzerland?

Motion Juillard 24.3209 calls for a legal basis that allows the Confederation to build and co-finance a sovereign digital infrastructure. Both chambers have adopted it. Motion Z'graggen 26.3221 calls for a stimulus programme and is with the National Council after the Council of States voted in favour. The Federal Council proposed rejecting both.

## Glossary

-   **Token**: The smallest billing unit of a language model; a word is roughly one to three tokens. Providers charge per million tokens, separately for input and output.
-   **Motion**: A parliamentary instrument in Switzerland that obliges the Federal Council to propose a law or measure once both chambers have approved it.
-   **e-ID**: The state-issued digital identity document of Switzerland, with which one can log in to authorities and services.
-   **Open weight**: Models whose weights are publicly available and may be self-hosted (e.g. Apertus, DeepSeek, Llama). A precondition for sovereign operation without vendor lock-in.
-   **Alignment**: Steering a model towards desired behaviour: which answers it gives, which it refuses and by which values it weighs them.
-   **Fine-tuning**: Further training of a pretrained model with selected examples so that it follows instructions or masters a specialised task. Also called post-training.
-   **Classifier**: A small model that assigns texts to a category, such as "sensitive topic" or "harmless". In training it serves as a filter or router.
-   **Pretraining**: The first and most expensive training phase, in which a model learns language and knowledge from very large amounts of text.
-   **Tokenizer**: The program that splits text into tokens before a language model processes it. Every model has its own, so the same text yields different token counts depending on the model.
-   **Gini coefficient**: A measure of inequality between 0 (all equal) and 1 (maximally unequal). For tokenizers it measures how differently many tokens the same content costs across languages.
-   **Vocabulary**: The fixed list of text pieces a tokenizer knows. Anything not in it as a whole is split into several tokens.
-   **CSCS**: The Swiss National Supercomputing Centre in Lugano, part of ETH Zurich. It operates the Alps supercomputer on which Apertus was trained.
-   **GPU**: Graphics processing unit. Language models run almost exclusively on GPUs because they execute very many computations in parallel.
-   **GPU hour**: One hour of compute on one graphics processor. The usual measure of compute effort when training large models.
-   **Parameters**: The learned numerical values of a language model. Their number describes its size: "8B" means eight billion parameters. Larger models can do more but need more GPU memory.
-   **Foundation model**: A generally trained language model on which other applications and adapted models are built.
-   **Personal data**: Any information relating to an identified or identifiable person, so customer records, HR files or even an email address. As soon as such data is involved, the revFADP applies.
-   **System prompt**: An instruction the operator invisibly places before every request. It sets the role, tone and limits of the model for an application.