Swiss AI Models Worth Knowing About

Swiss AI has quietly built a small but serious catalogue of models, and almost all of it follows the same idea: keep it open, keep it auditable, and make it work across languages. Rather than chasing the biggest chatbot on the market, universities and research institutes have focused on models that organisations can inspect, host themselves and adapt to their own needs. Here are the ones that matter right now.

The Models at a Glance

Apertus 8B

The smaller of the two original Apertus models is the one most developers will actually run, since it fits on far more modest hardware.

Features

• Trained on 15 trillion tokens covering more than 1,000 languages, with roughly 40% of the data in languages other than English
• Includes underrepresented languages such as Swiss German and Romansh
• Available as both a base model and a chat-tuned instruct version
• Works with Transformers, vLLM, SGLang, llama.cpp and MLX
• Context window of around 65,000 tokens

Strengths

• Cheap to host, which suits startups and small teams
• Fully open weights, data recipes and training documentation
• Easy to fine-tune for a specific language or industry

Apertus 70B

This is the flagship, and the model Switzerland points to when it talks about sovereign AI.

Features

• Same training corpus and multilingual coverage as the 8B, with far more capacity
• Trained on the Alps supercomputer in Lugano using thousands of GH200 GPUs
• Uses the Goldfish training objective to reduce verbatim memorisation of training text
• Respects robots.txt opt-outs, including retroactively, and filters personal data from its training set
• Deployable through Swisscom's sovereign platform and Amazon SageMaker JumpStart

Strengths

• Built with Swiss data protection, Swiss copyright law and the EU AI Act's transparency rules in mind, which matters for regulated organisations
• Permissive licence that allows commercial use
• Complete transparency, including intermediate checkpoints, so auditors can see how it was built

Apertus 1.5

Released in July 2026, this update takes feedback from the first release and adds new capabilities instead of starting from scratch. The team has said regular updates will follow over the coming years.

Features

• Refinements based on real-world use of the first Apertus models
• A dedicated inference service at CSCS for its users and the wider Swiss academic community
• Full model suite published on Hugging Face with documentation on the project site

Strengths

• Signals a long-term commitment rather than a one-off release
• Gives universities and researchers direct access without building their own infrastructure
• Keeps the same open approach, so earlier work and fine-tunes remain relevant

MeditronFO

Medicine is where Apertus is starting to branch out. MeditronFO is described as the first fully open framework for building medical language models, and EPFL researchers use Apertus as one of its foundation models.

Features

• A framework for creating medical LLMs rather than a single fixed chatbot
• Built on open foundation models, including Apertus
• Reported testing at Lausanne University Hospital (CHUV) since May 2026

Strengths

• Openness lets hospitals and researchers inspect and adapt the model, which is hard to do with closed medical AI
• Multilingual roots suit a country where clinicians work in German, French and Italian
• Grounded in academic medical research rather than a commercial product roadmap

SwissBERT

SwissBERT is older and far smaller than Apertus, but it still does a job the big models handle less precisely. It is a masked language model, meaning it is built to understand and classify text rather than write it.

Features

• Trained on more than 21 million Swiss news articles, about 12 billion tokens
• Language adapters for Swiss Standard German, French, Italian and Romansh Grischun
• A vocabulary tuned to Switzerland-related text
• Built on the X-MOD architecture

Strengths

• Outperforms earlier models on Switzerland-related tasks, especially with contemporary news and Romansh
• Lightweight enough for classification, entity recognition and text alignment jobs
• Can be extended with further adapters, such as Swiss German dialects

Limitations to keep in mind

The training data is news writing up to 2022, so performance can drop on other domains. The model licence is non-commercial, which rules it out for most business projects.

Where Swiss Models Still Fall Short

Supporting more than a thousand languages on paper is not the same as performing well in all of them. Some testers have found Swiss German and Romansh output weak, and dialect handling remains the hardest part of the puzzle. Apertus also trades a little raw polish for transparency, so teams chasing top benchmark scores may still look elsewhere.

Picking the Right One

For general-purpose work with the option to self-host, Apertus 70B is the natural choice, with the 8B version covering lighter needs. Academic users should look at Apertus 1.5 through the CSCS service. Healthcare projects should follow MeditronFO, and anyone doing Swiss-focused text analysis on a non-commercial basis will get good value from SwissBERT.

› More Article