fairlane.systems

CH MULTILINGUALISM · COMPLIANCE

AI and Swiss multilingualism: LLMs for German, French, Italian and Romansh

Swiss language split (2025): German 62.8%, French 22.9%, Italian 8.2%, Romansh 0.5%. Which models handle the four national languages plus Swiss German – as of May 2026.

Researched & fact-checked by: · As of: 2026-05

What does "CH-capable" mean for an LLM?

Switzerland has four official national languages. The 2025 split per the Federal Statistical Office: German 62.8 percent, French 22.9 percent, Italian 8.2 percent, Romansh 0.5 percent. English is not an official language but is increasingly used in business (around 45 percent of the working population report regular use).

For a language model, "CH-capable" means more than understanding German. It covers Standard German in its Swiss variant (helvetisms: Velo, Trottoir, Coiffeur, parkieren, Treuhand), Swiss proper nouns (cantons, municipalities, cantonal authorities with genitive constructions), the dialect-Standard-German switch in business contexts, and legal jargon from Swiss sources (Fedlex, FDPIC, FINMA). In French, the distinction between France French and Swiss Romand matters (huitante instead of quatre-vingts in some cantons, the dejeuner shift). In Italian, Ticino Italian with Roman loanforms and administrative terms counts.

Written Swiss German is a separate question. There is no standardised orthography – spellings vary by canton and generation. A model that answers email must accept "Gruezi", "Hoi zaeme", "merci villmal" as valid greetings without translating or "correcting" them. Audio is harder still: speech recognition (Whisper large-v3, Apertus Audio) transcribes Bernese, Zurich and Wallis dialects with different quality.

Why CH multilingualism matters for AI projects

Three reasons make the language question a risk factor – not a courtesy.

First: client effect. A fiduciary office in Geneva, Lugano or Chur cannot force clients to communicate in German. If AI-assisted client replies only work in Standard German, the office fails in Romandie and Ticino on day one. Swiss fiduciary associations (TREUHAND SUISSE, EXPERTsuisse, fiduciaire|suisse) are organised trilingually – supporting tools must be too.

Second: legal duties in several sectors. Authority communication and contracts in multilingual Switzerland follow language rules (Languages Act SR 441.1, Languages Ordinance SR 441.11). Banks and insurers must provide customer-relevant documents in the client language, which forces model choice with FR/IT quality for automated correspondence. In criminal law, Federal Constitution Art. 31 para. 2 requires information in an understood language.

Third: quality is not symmetric. Even top models (the current top GPT model, the current top Claude model, Gemini 2.5) are markedly stronger in German than in Italian or in Swiss Romand. Picking a model purely on English benchmarks (MMLU, HumanEval) risks production error rates 3-5x above expectation in Swiss use. Multilingual benchmarks (MGSM, XNLI, Belebele) are indispensable.

Model comparison for Swiss languages as of May 2026

The following overview is based on MGSM scores, in-house evaluation runs and public benchmark tables (Hugging Face Open LLM Leaderboard, MMLU-Pro multilingual, SwissLegalBench).

German (Swiss Standard German): Very good: the current top GPT model, the current top Claude model, Gemini 2.5, Mistral Large 2, Apertus 70B. Workable: Llama 4 70B, Command R+, DeepSeek V3. Weak: small 7B models without German training share. Helvetism recognition is best in the current top Claude model and Apertus, since these were trained with Swiss corpora (federal council minutes, Tages-Anzeiger, NZZ archive under licence).

French (with Swiss Romand): Top: Mistral Large 2 (Mistral AI, Paris – best FR/DE model), the current top Claude model, the current top GPT model. Apertus 70B is on par with Mistral for FR, because EPFL/ETH Zurich included Romandie corpora. Swiss proper nouns (cantons, communes vaudoises) are most accurate with Apertus and Mistral.

Italian (with Ticino Italian): Solid: the current top GPT model, the current top Claude model, Gemini 2.5, Mistral Large 2. Italian is the weakest of the three dominant Swiss languages in practically all models – training data share is 5-8x smaller than for German or French. Ticino Italian is often interpreted as Tuscan Italian; local administrative terms (Comune politico, Patriziato, Cancelleria dello Stato) are not reliably captured.

Romansh (Sursilvan, Sutsilvan, Surmiran, Puter, Vallader, Rumantsch Grischun): Only Apertus 70B delivers acceptable quality, because ETH Zurich integrated RM corpora from Lia Rumantscha. The current top GPT model and the current top Claude model recognise RM but average 30-50 percent inflection errors. Practical recommendation May 2026: route RM workflows through Apertus with human post-editing.

Swiss German (audio): Whisper large-v3 recognises Bernese and Zurich dialect at a word error rate (WER) of 18-25 percent, Wallis and Grisons dialect at 35-45 percent. Apertus Voice (in development May 2026, not yet in production) targets RM and CH-DE dialects but is not public. For production phone transcription, Whisper plus re-editing is the current state of practice.

Check CH multilingualism in 6 steps

  1. 01Map client and use-case languages: what DE/FR/IT/RM/EN shares appear in production?
  2. 02Benchmark candidate models on MGSM and on 50-100 real client texts per language (synthetic or anonymised).
  3. 03Build a helvetism list: 30-50 typical Swiss terms (Velo, Treuhand, Bezirksgericht, AHV number) as a sanity check.
  4. 04Define routing: primary model per language (Mistral Large 2 for FR, Apertus for RM/CH-DE, das aktuelle Claude-Spitzenmodell for DE/IT) plus fallback.
  5. 05Audio pipeline: Whisper large-v3 for phone transcription, tracking word error rate by dialect region.
  6. 06Periodic re-test every 6 months, because new model releases (Apertus 2, an upcoming Mistral Large generation, the current Claude model) reshuffle the ranking.

When multilingual model choice is mandatory

A multilingual check is mandatory in four cases.

First: clients in Romandie or Ticino. Anyone serving clients in the cantons Geneva, Vaud, Neuchatel, Jura, Fribourg, Wallis, Ticino or Grisons must test FR or IT capability. A routing setup that dispatches by client language to Mistral Large 2 (FR), Apertus 70B (RM/CH-DE) or the current top Claude model (DE/IT) is standard at Swiss consulting boutiques in May 2026.

Second: authority communication and contract drafting. Contract drafts and correspondence with Swiss authorities must be written in the applicable official language – wrong conjugation or wrong terms (e.g. "Notar" vs "notaire" vs "notaio" with cantonal differences) is noticed immediately and erodes trust.

Third: phone and voicebot setups. A Swiss hotline must serve DE/FR/IT at minimum. RM capability is a plus, rarely a must (limited speaker base).

Fourth: audit and accounting workflows. Supplier invoices arrive in all four national languages plus English. Receipt OCR with downstream LLM classification must handle all languages – small models often fail on Italian or RM.

When a monolingual solution suffices

Three patterns can legitimately stay monolingual.

A purely German-Swiss office with clients exclusively from DE cantons (ZH, BE-centre, AG, BS, BL, LU, OW, NW, UR, SZ, ZG, SO, SH, TG, AR, AI, SG, GL): FR/IT capability is a reserve, not a production path. A model with strong German suffices.

Internal tools without client contact – e.g. code generation, internal research, technical documentation in English: English quality counts, Swiss multilingualism is secondary.

Proof-of-concept phase: a PoC in German is legitimate as long as the next phase (pilot with real clients) has an FR/IT extension on the roadmap.

This is not legal advice. Language requirements from sectoral regulation (FINMA, ESTV, KESB) require qualified legal counsel – the points named here are not exhaustive.

Trade-offs

STRENGTHS

  • Apertus covers RM and CH-DE dialects better than any cloud model
  • Mistral Large 2 delivers top FR quality with an EU hosting option
  • Multi-provider routing allows the best model per language
  • Clear language compliance (Languages Act, client-language right) is achievable

WEAKNESSES

  • Italian remains the weakest of the three dominant Swiss languages across all models
  • Swiss German audio has 18-45 percent WER – not verbatim-grade
  • Multi-provider routing increases operational complexity and compliance burden
  • Apertus Voice not yet productive – RM audio remains weak in May 2026

FAQ

Which model is best for Swiss French as of May 2026?

Mistral Large 2 from Mistral AI (Paris) delivers the strongest results on Romand texts. Apertus 70B is on par, with advantages on Swiss proper nouns and cantonal administrative terms. The current top Claude model is solid, the current top GPT model slightly behind. For EU hosting needs: Mistral Large 2 via Microsoft Azure Switzerland North region or via Infomaniak.

Can Whisper transcribe Swiss German?

Limited. Whisper large-v3 transcribes Bernese and Zurich dialect at WER 18-25 percent, Wallis/Grisons dialect at 35-45 percent. Enough for keyword detection, not verbatim minutes. Recommendation: Whisper plus human post-editing. Pure CH-DE specialty models (Apertus Voice, in development May 2026) are not yet productively available.

Do I need a separate model for Italian?

Rarely. The current top Claude model, the current top GPT model, Mistral Large 2 and Gemini 2.5 cover Italian acceptably. The risk is at Ticino administrative terms (Comune politico, Patriziato, Consiglio di Stato) and cantonal proper nouns – a glossary file in the RAG pipeline that enforces such terms as mandatory vocabulary helps here.

How many clients speak Romansh at all?

About 40-60 thousand people, primarily in Grisons. In canton GR, RM is an official language alongside DE and IT. In practice a fiduciary office in Chur or the Engadin should at least be able to read RM correspondence – Apertus 70B handles that – and should engage an RM editor (Lia Rumantscha) for official documents. Pure RM generation without human final review is not production-ready in May 2026.

Related topics

APERTUS · COMPLIANCEApertus: the open Swiss AI model from ETH Zurich, EPFL and CSCS – status May 2026MISTRAL · LLM PROVIDERMistral AI from a Swiss fiduciary perspective: EU residency, pricing, sovereigntyROUTING · AI CONCEPTMulti-LLM routing: which model when, for how muchrevDSG · COMPLIANCErevDSG / revFADP and AI: what the revised Swiss Data Protection Act means for LLM useVOICE · SERVICEVoice agent on the phone: AI that calls and is called

Sources

  1. Bundesamt für Statistik BFS – Sprachen in der Schweiz (Erhebung 2025) · 2026-05
  2. Sprachengesetz (SpG, SR 441.1) – Fedlex Volltext · 2026-05
  3. Apertus Modell-Card ETH Zurich/EPFL/CSCS · 2026-05
  4. Mistral AI – Mistral Large 2 Modell-Card · 2026-05
  5. OpenAI Whisper large-v3 Modell-Card · 2026-05
  6. Lia Rumantscha – RM-Korpora und Sprach-Standards · 2026-05

FITS YOUR STACK?

This is not legal advice. We assess model choice for DE/FR/IT/RM clients, build routing across Apertus, Mistral and Claude, and benchmark audio pipelines for Swiss dialects. Initial call free of charge.

Book a call