Apertus: Switzerland's Open Language Model, Explained
What EPFL, ETH Zurich and the Swiss National Supercomputing Centre CSCS have built, what Apertus can do and where it still falls short. And how to use it in your organisation today, without buying your own graphics cards.
What is Apertus?
Apertus is a fully open, multilingual language model from Switzerland. It was developed by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) as part of the Swiss AI Initiative, and released on 2 September 2025 in two sizes with 8 and 70 billion parameters.
"Open" means more with Apertus than with most other models: not only the model weights are freely available, but also the training data, the training recipes and intermediate checkpoints. The Apache 2.0 licence allows commercial use as well.
Apertus is a large language model like the models behind ChatGPT. It is not a ready-made chat service, though, but a building block: you use it through a platform or an API, or run it yourself. Read more about the principle of open models in the glossary under open-source AI.
Apertus in numbers
Figures from the model card on Hugging Face and from CSCS, as of October 2026.
Apertus 1.5: what the new version can do
On 24 July 2026, CSCS released Apertus 1.5. The new version builds on 1.0 with further training: 4 trillion additional tokens for the 8B model and 2 trillion for the 70B model. Licence and openness stay the same.
Understands images
Apertus 1.5 accepts images as well as text. Audio input is included as an experimental feature.
Thinking mode
An optional thinking mode lets the model work through intermediate steps before it answers. This helps with multi-step tasks.
Four times the context
According to Public AI, the context window grows to around 262,000 tokens. Long contracts or entire dossiers fit into a single request.
Better with tools
Instructions are followed more precisely, and the model is better at calling tools and APIs. That is the foundation for AI agents.
Apertus Mini
In addition, there are 16 compact, smaller variants. They run on weaker hardware, such as laptops or devices.
No benchmarks yet
Public AI has announced a technical report with measurements for version 1.5 in the coming weeks. Until then, only the figures for version 1.0 are available.
Apertus compared with Llama, Qwen and ChatGPT
The developers compared the base models of Apertus 1.0 with open models of a similar size. The score is the average of six common tests for general knowledge, reasoning and multilinguality (ARC, HellaSwag, WinoGrande, XNLI, XCOPA, PIQA).
| Model | Size | Average |
|---|---|---|
| Apertus 70B | 70 bn | 67.5 |
| Llama 3.1 70B (Meta) | 70 bn | 67.3 |
| Qwen2.5 72B (Alibaba) | 72 bn | 69.8 |
| Apertus 8B | 8 bn | 65.8 |
| Llama 3.1 8B (Meta) | 8 bn | 65.4 |
| Qwen2.5 7B (Alibaba) | 7 bn | 64.4 |
Source: Apertus model card on Hugging Face, base models version 1.0.
What the numbers say: Apertus is roughly on par with Meta's Llama 3.1 and slightly behind Qwen2.5. For a model whose training data is fully disclosed and which respects opt-out requests from data owners, that is a good result.
What they don't say: Apertus was not compared directly with the largest commercial models such as GPT, Claude or Gemini. For complex tasks, long texts and programming, these remain clearly stronger. Apertus stands out for transparency and control, not for top performance.
Six ways to use Apertus
From a free test in the browser to running it in your own data centre. Which way fits depends on whether you just want to try it or work with company data.
-
1
Public AI: test it in the browser
The non-profit initiative Public AI offers a chat with Apertus: the 8B model without sign-up, the 70B model with thinking mode after login. Good for trying it out, not meant for confidential data.
-
2
Hugging Face: download the model
The weights are freely available on Hugging Face, for example under
swiss-ai/Apertus-70B-Instruct-2509and in the collection for version 1.5. For developers who want to run or adapt the model themselves. -
3
Locally on a laptop
In a compressed (quantised) form, the 8B model also runs on a laptop, for example on Apple computers with Apple Silicon. According to Liip, a graphics card with 8 GB of memory is even enough.
-
4
Through an API
Swisscom has offered Apertus to business customers on its Swiss AI Platform since September 2025. Infomaniak offers Apertus 1.5 with 70 billion parameters via API. Useful if you build your own applications.
-
5
In the Evoya platform
In Evoya, Apertus is available alongside more than 30 other models, with no installation. Your team uses it in the chat, in AI agents and with its own documents, managed by your organisation.
-
6
In your own data centre
The 70B model needs server graphics cards. As a rough guide (our own estimate): around 140 GB of GPU memory at full precision, around 40 GB in compressed form. We plan and run this for you: AI on-premise.
What Apertus is good for, and what it isn't
Our assessment from practice. Test with your own examples before you decide.
A good fit
- Summarising, translating and rewording texts
- Answering questions about internal documents (RAG)
- Sorting emails and requests and extracting data
- Applications where it must be traceable what the model was trained on
- Public administration and education, where open Swiss technology is preferred
Better use another model
- Complex analyses over many steps
- Programming and technical troubleshooting
- Demanding texts where every phrase has to be right
- Tasks that need current knowledge from the internet
Our advice
- Don't pick one model for everything
- Process sensitive data with Apertus or other Swiss models
- Allow stronger international models for non-critical tasks
- Decide per use case which model is allowed
Why a Swiss model alone doesn't guarantee data protection
Apertus comes from Switzerland. But where your data is processed is not decided by where the model comes from. It depends on where it runs and who operates the server. A Swiss model on an American cloud service is, legally, an American service.
Three questions help you choose: Where is the server? Who has access to inputs and outputs? And what does the contract say, for instance about storage and whether your data is used for AI training?
At Evoya, the open-source models run in Switzerland, and your data is never used for AI training. If data must not leave your network at all, we run Apertus as private AI in a dedicated Swiss environment or on-premise at your site.
Using Apertus in Evoya
Evoya's Swiss AI platform brings Apertus and more than 30 other models together in one interface that your organisation manages itself.
Hosted in Switzerland
Open-source models such as Apertus run on Swiss infrastructure. Your inputs are never used for AI training.
Choose a model per task
Apertus for sensitive data, ChatGPT or Claude for non-critical research. Admins decide which models your team may use.
With your knowledge and agents
Connect Apertus to your own documents and build AI agents that complete tasks on their own.
All features free for 14 days, no credit card required. After that from CHF 29 per month; all prices are on the pricing page.
Frequently asked questions about Apertus
Yes, the model itself is free. It is released under the Apache 2.0 licence, and you can download the weights freely. Costs arise for running it: either for your own servers with graphics cards or for a service that runs Apertus for you.
No, not in raw performance. In the developers' tests, Apertus is roughly on par with Meta's Llama 3.1; the largest commercial models such as GPT are stronger at complex tasks. The strengths of Apertus are transparency, open training data and the option to run it entirely in Switzerland.
To try it out, in the Public AI chat. As a download on Hugging Face. Via API at Swisscom or Infomaniak. And without installation in the Evoya platform, alongside more than 30 other models. For company data, we recommend a service with Swiss hosting and a clear contract.
According to ETH Zurich, Apertus also covers Swiss German and Romansh. Quality for dialects and smaller languages is lower than for Standard German, French or English, though. For business texts, we recommend Standard German and a test with your own examples.
Yes, and more open than most models. Besides the model weights, the training data, the training recipes and intermediate checkpoints are published too. Many other "open" models only disclose the weights.
EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS), as part of the Swiss AI Initiative. The model was trained on CSCS's Alps supercomputer.
Yes. The Apache 2.0 licence allows use, adaptation and redistribution in commercial products too. Observe the licence terms, for example the licence notice when redistributing.
Apertus 1.5 was released on 24 July 2026. It understands images as well as text, has an optional thinking mode, a context window around four times larger and works better with tools. In addition, there are 16 compact variants under the name Apertus Mini.
Sources (retrieved 31.10.2026): Apertus model card on Hugging Face · CSCS, Apertus 1.5 · Public AI, Apertus 1.5 · Swisscom, Apertus on the Swiss AI Platform · Liip, six ways to try Apertus 1.5. The assessments of fit and the hardware guide are our own.
Use Apertus with your team
Try Apertus and more than 30 other models free for 14 days. Or talk to us about how to use Apertus safely with your data.