July 31, 2026
Apertus 1.5 for Swiss Companies: What the Open Swiss Model Solves — and What You Have to Deliver Yourself
Apertus 1.5 from ETH, EPFL and CSCS was released on 24 July 2026. We assess which use cases the Swiss model suits, the three ways to obtain it, and the operational duties its acceptable use policy creates.
Demand for a "Swiss solution" is real. Since 24 July 2026 there is a serious basis for one — but it answers only half the question clients actually ask.
On that date, ETH Zurich, EPFL and the Swiss National Supercomputing Centre CSCS released Apertus 1.5. It has come up in nearly every advisory conversation since. This article answers the three questions that recur: what the model does, how you obtain it, and what you have to do yourself.
What Apertus 1.5 is technically
Apertus is developed within the Swiss AI Initiative and trained on the CSCS "Alps" supercomputer in Lugano. Version 1.5 comes in two sizes, 8 and 70 billion parameters, accompanied by a set of compact models ("Apertus Mini") distilled from the 8B model — 16 variants for memory-constrained environments. Three things are new compared with the September 2025 release: - Multimodality: the model takes images and audio alongside text as input, and generates text. - Thinking mode: it can work through tasks step by step before answering. - Context window: 262,144 tokens, four times the previous version — very large document sets in a single request. What sets it apart from other open models such as Llama, Mistral or Qwen is not size but disclosure: not only the weights are published, but the data pipeline and training recipes too. The model's provenance is auditable — for most "open-weight" models it is not.
The confusion that shapes almost every conversation
Anyone asking for a Swiss AI solution generally wants to know where their data is processed. That is precisely the question a model does not answer. A model is a file. It has no location and processes nothing until somebody operates it. The place of processing is created by the operator — a decision entirely independent of model choice. Two consequences that surprise people in practice: You can run Apertus on AWS SageMaker or Microsoft Azure; the official provider page lists both. A Swiss model then runs on US infrastructure, and the situation you set out to avoid persists unchanged. Conversely, a commercial model such as Claude or GPT under a proper data processing agreement with European processing may be the cleaner construction in data-protection terms than a poorly configured self-hosted Apertus. The relevant question is therefore not "Swiss model or not?" but: who processes, under which law, under which contract? What a Swiss model adds is nevertheless substantial — it is simply something else: independence from a single vendor's licensing and pricing decisions, operability over many years, and the option to adapt the model to your own data without handing that data over.
Three ways to obtain it — with very different effort
Route 1: Managed API from Switzerland. Swisscom offers Apertus via its Swiss AI Platform, Infomaniak as a pay-per-token interface, Phoenix Technologies as model-as-a-service; CSCS runs its own inference service. The effort equals a normal API integration, processing stays in Switzerland, and the provider handles operations. For the large majority of use cases this is the right entry point. Negotiate terms directly with the provider — we could not verify published price lists consistently across the Swiss offerings and therefore quote none here. Route 2: Public testing. Apertus is freely accessible through the Public AI platform. Suitable for trying it out and forming a first impression of quality, not for company data — the usual caveats for any public service apply. Route 3: Self-hosting. Full control, in your own data centre or at a Swiss host. Operations are the real cost block — not the hardware, but staff, model maintenance and on-call readiness. We worked through when this pays off in Local LLMs for SMEs; that threshold logic applies to Apertus unchanged.
Where we consider Apertus a good fit
- Government-adjacent and regulated environments where model provenance forms part of the justification. The Canton of Ticino uses Apertus to translate official documents. - Domain-specific further development. EPFL uses Apertus as the basis for its medical model Meditron. Adapting a model to your own domain language requires open weights — a closed API cannot do it. - Multilingual tasks in a national context, particularly across the Swiss national languages. - Long-lived applications. Nobody can switch off or reprice an Apache-2.0 model on you. With commercial vendors, the retirement of a model version is a genuine operational risk. - Editorial applications with a transparency claim. The Basel outlet Bajour uses Apertus in its newsroom for political analysis.
What Apertus does not solve — four points that belong before the decision
1. Apache 2.0 does not mean "no obligations." The weights are Apache 2.0, commercial use included. The Hugging Face repositories are gated, however: registration, contact details and acceptance of an acceptable use policy are required. Three things follow from that policy — you process personal data as an independent controller, you indemnify ETH Zurich and EPFL against third-party claims arising from your use, and deletion requests concerning personal data held in the model are handled via a hash-value file that you apply yourself as an output filter. The recommendation is to refresh it every six months. That is not an installation step but a standing operational process, with an owner and a date. 2. Robust comparative figures are currently missing. The model cards defer benchmarks to a technical report that has not yet been published. Anyone claiming today that Apertus 1.5 leads or trails GPT-5.6, Claude Sonnet 5 or Gemini 3.1 Pro cannot support that with published numbers. The only assessment that holds right now is the one run on your own tasks. 3. The "1000 languages" is a training figure. The project site cites over 1000 languages and the FAQ more than 1800 — yet the same FAQ states the model is fully conversational in a few dozen. What matters for you are the national languages and the quality of Swiss Standard German, and you establish that on your own texts, not from a language list. 4. Model maintenance becomes your job the moment you self-host. Ten months separate the September 2025 release from version 1.5. Every version change needs evaluating, testing and rolling out. In a managed API that effort is included in the price; in self-hosting it is not.
The decision logic in four questions
Is the place of processing mandated contractually or by regulation? If not, Apertus is one option among several and has to win on merit. If yes, only Swiss operating routes qualify — with Apertus or another open model. Do you need to adapt the model to your own data? If yes, you need open weights. This is the strongest substantive argument for Apertus. Is auditable model provenance part of your justification to third parties — regulators, clients, the public? Then the disclosed data pipeline is an argument no commercial vendor can match. Do you have someone to own the operational duties arising from the acceptable use policy? If not, the route runs through a managed API, where some of those duties can be addressed contractually.
Our recommendation
For most Swiss SMEs the pragmatic route is this: start with a managed API from Switzerland and evaluate Apertus on three to five real tasks from your operations — not on demo examples. Only once it passes that test does the question of adaptation or self-hosting arise. What we explicitly advise against: using Apertus as a data-protection argument without having settled the processing route. The model's origin is a good argument — it simply is not the thing that protects your data. We describe how to approach the data-protection side systematically in nDSG compliance for AI systems; the questions to put to a vendor are in our AI vendor selection specification. We support this evaluation as an engagement: requirements, choice of route, a test setup on your own tasks, and assessment of the outcome. Get in touch. This article reflects the position as of 31 July 2026 and draws on publications by the ETH AI Center, the Hugging Face model cards, the Swiss National AI Institute's Acceptable Use Policy, and the information on apertus-ai.org. It does not constitute legal advice.