In September 2025, Switzerland released its own language model. It generated a lot of attention—and raised many questions: Where can I try out Apertus? Is it as good as ChatGPT? And is it worth it for my company? A year later, with version 1.5 and a growing number of providers, now is a good time to take a sober look at these questions.

First, the short answer: The easiest way to try Apertus for free today is in Proton Lumo or in Euria from Infomaniak. For businesses, Swisscom, Infomaniak, and other Swiss providers offer Apertus via an API. And when it comes to complex tasks, the major commercial models still lead the way.

What is Apertus?

Apertus is the fully open-source, multilingual language model developed by the Swiss AI Initiative, a collaboration between ETH Zurich, EPFL, and the Swiss National Supercomputing Center (CSCS). It was trained on the Alps supercomputer in Lugano. The first version was released on September 2, 2025, in two sizes: one with 8 billion parameters and one with 70 billion parameters.

At Apertus, “open” means more than it does with most models. With many “open” models, you can only download the final weights—that is, the trained model itself. With Apertus, training data, code, methods, and intermediate results are also open. To illustrate: Most vendors sell you the finished cake, but Apertus also puts the recipe, the ingredient list, and the baking times on the table.

Key figures according to the technical report:

  • Training dataset size: 15 trillion tokens, of which about 40 percent are in languages other than English.
  • Languages: Trained on over 1,800 languages, including all four national languages as well as Swiss German. According to the official FAQ, Apertus is truly capable of conversing in several dozen languages.
  • License: Apache 2.0, which means it can also be used for commercial purposes. To access the weights for version 1.5, you must also agree to the terms of use on Hugging Face; access is granted automatically.

What's new in Apertus 1.5?

Apertus 1.5 was released on July 24, 2026. It is not a new model, but rather an update: The 8B variant was further trained with 4 trillion additional tokens, and the 70B variant with 2 trillion additional tokens—for the first time, using image data as well.

  • Much longer context: up to 262,144 tokens—four times as much as before. This allows you to process long documents in a single step.
  • Thinking Mode: If enabled, the model will think through its response step by step before answering. This feature is turned off by default. According to a study by the University of St. Gallen, it significantly improves performance in math but reduces adherence to instructions.
  • Image Understanding: Apertus can read and describe images; audio input is possible on an experimental basis. The output is always text.
  • Improved instruction adherence and tool utilization: important when integrating Apertus into workflows or agents. Compared to version 1.0, the 8B variant showed double-digit gains in standard tests.

The information remains publicly available: weights, data, and interim results for Version 1.5 are also public; the detailed technical report will follow.

Why Apertus?

What's the point of a Swiss model when there's already ChatGPT, Claude, and Gemini? The answer comes down to three points.

Traceability. Because the data and training are publicly available, it is possible to verify what the model was trained on. For government agencies, hospitals, or companies with strict procurement and auditing requirements, this makes a real difference.

Control over your data and operations. You can run Apertus yourself or use it through a Swiss provider. Your data doesn't have to leave the country, and you aren't at the mercy of the prices or rules of a single U.S. corporation.

Languages of Switzerland. The national languages and Swiss German were the stated training goals. In practice, however, weaknesses still appear in written Swiss German: The translation service Helvetra reports that Apertus does not reliably distinguish between dialect and Standard German. There are still no reliable tests for spoken dialect.

It’s also important to note what Apertus is not. According to a swissinfo analysis, the 70B variant lags significantly behind open models of the same size in programming, mathematics, and logic. Parliament did not choose Apertus for its assistant “Pia” because, according to SRF, it tends to make things up. Apertus, however, excels at answering questions in German and at translation: In the MÖVE administrative benchmark conducted by the German Federal Printing Office, the 70B variant of version 1.0 ranked 3rd out of 39 models in answering questions and 10th in summarizing.

In a nutshell: Apertus stands out for its openness and data sovereignty, not for its top performance. It’s sufficient for many everyday tasks, but not for demanding analyses.

Here's how to try out Apertus

You don't need to install anything to get a first look. Since fall 2026, several chat services have included Apertus directly in their menu.

  1. Proton Lumo, free: Starting September 17, 2026, you can select the Apertus 1.5 AI assistant from the model menu in Proton Lumo—according to Proton, this option is available to all Lumo users, including those on the free plan. Proton’s own model remains the default; you must actively select Apertus. Proton operates Lumo on its own infrastructure in Europe.
  2. Infomaniak Euria, free: In the Swiss Euria assistant, you can select Apertus in the chat. Euria is free with a daily data allowance, and the data remains in Switzerland.
  3. Public AI: The 8B version is available at publicai.co/chat without requiring registration, while the 70B version is available at chat.publicai.co with a free account and Thinking Mode. According to Public AI, computations are performed by partners worldwide, and data can be shared with researchers. This is not the right place for confidential content.
  4. mono.chat: The service has been offering Apertus 1.5 70B on all paid plans since August 2026, starting at $20 per month. Customer data is stored in Frankfurt; mono does not disclose which provider hosts the platform.

A comparison of Swiss AI chatbots shows which chat service uses which models and where the data is stored.

Here's an example to get you started: Take a public document—such as your company's terms and conditions—and have Apertus and your usual chat app answer the same three questions about it. In just five minutes, you'll see where the differences lie that are relevant to your work.

Apertus in the Workplace: Interfaces and Hosting

Anyone who wants to integrate Apertus into their own applications will need a programming interface. There are now several Swiss providers offering such interfaces. The Swiss AI Initiative itself maintains the official list of providers.

  • Swisscom: The Swiss AI Platform offers Apertus 1.5 70B for business customers, billed per token. Swisscom does not support Thinking Mode for tool usage.
  • Infomaniak: The AI tools include Apertus 1.5 70B as a beta version, with free credit to get you started and data hosted in Switzerland. Up to 100,000 tokens can be entered per request.
  • PHOENIQS (kvant.cloud): Apertus 1.5 70B with full context and published prices per million tokens, geared toward businesses and the public sector.
  • stepping stone: The Bern-based provider offers the 8B plan as a service billed on a pay-as-you-go basis.
  • Safe Swiss Cloud: Apertus 1.5 70B as part of the “Private AI” offering, with data stored in Switzerland.
  • Universities: Researchers at Swiss universities use the CSCS inference service.

Three questions can help you make your decision: Where is the data processed? Which version and size are running there? And how much does a typical request cost compared to your current model? Liip calculated for its city chatbot that Apertus 70B at Infomaniak was about four times more expensive per token than a comparable commercial model.

Running Apertus Locally: What Do You Need?

A local model is useful when data must not leave the device at all or when you want to work without an internet connection.

  • 8B variant: According to the official overview, it requires about 6 GB of RAM in a scaled-down version and therefore runs on many laptops.
  • 70B variant: In the standard 4-bit version, it requires approximately 48 GB of memory; at full precision, it requires many times that amount. This requires server-grade hardware or a very well-equipped computer.
  • Apertus Mini: A smaller version even runs directly in the browser, with no installation required.

The weights are available on Hugging Face under "swiss-ai." LM Studio and Ollama offer community-contributed versions, though Apertus is not officially supported on Ollama. For version 1.5, these versions are currently text-only models without images or audio. To see what runs on your hardware and how Apertus compares to other open-source models, check out the calculator on the "AI Models Compared" page.

Where Apertus Is Already in Use

According to swissinfo, after one year, the Swiss AI Initiative has recorded over 4 million downloads and at least 70 external deployments. Here are a few examples from Switzerland:

  • Canton of Ticino: The cantonal IT center runs a finely tuned Apertus 8B in its own data center. About 100 employees are testing it.
  • Medicine: EPFL has introduced MeditronFO, an open framework for medical language models based on Apertus.
  • Media: The Basel-based online magazine Bajour uses a locally hosted Apertus instance for its newsletter and to analyze Grand Council minutes.
  • Community: Hack Apertus, a series of hackathons running from October 1, 2026, through May 2027 with over 120 partner organizations, is now underway. Learn more on the event page.

When is Apertus a good fit for businesses?

Apertus is not a substitute for every AI tool. It really shines in certain situations:

  • Sensitive data— such as personnel files, patient records, or confidential contracts—that should not be stored in a foreign cloud should be hosted by a Swiss provider or managed locally.
  • Government and Procurement: Where Transparency Regarding Origin and Training Is Required.
  • Translating between national languages: In a legal test, Apertus 1.5 translated 70B court decisions better than all 16 other models tested, but performed poorly when it came to summarizing.
  • Custom applications: Chatbots that respond based on your own documents—so-called RAG applications. The model first searches your documents and then formulates the response. It’s worth testing this with real questions before you make a decision.

For demanding programming or analysis tasks, you're better off choosing one of the larger models, ideally from a provider with clear contract terms.

What's Next

According to SRF, training for Apertus 2.0 is scheduled to be completed in December 2026, and the model will be many times larger. The initiative’s roadmap calls for Version 2.0 to be released in the first quarter of 2027. It remains to be seen whether Apertus will catch up with the larger models, but the direction is clear.

swissAI and Apertus

swissAI is an independent association and is not part of the Swiss AI Initiative from an organizational standpoint. However, we actively support its goals through knowledge transfer to the field, networking, and practical applications. Dr. Imanol Schlag, co-lead of Apertus, serves on the swissAI Advisory Board. Learn more on the Swiss AI Initiative and Apertus page.

Frequently Asked Questions

Is Apertus free?

The model itself, yes. It is licensed under the Apache 2.0 license and may be freely used, modified, and distributed, even for commercial purposes. You can try it for free in Proton Lumo, Infomaniak Euria, and Public AI. Costs apply if you use it via an API or host it yourself.

Where can I try out Apertus?

The easiest way is through the model menu in Proton Lumo or Infomaniak Euria—both are free. If you don’t have an account, you can use the 8B version at publicai.co/chat. For confidential data, a Swiss provider like Infomaniak or a local company is a good choice.

Does Apertus speak Swiss German?

Swiss German is one of the more than 1,800 languages used to train Apertus. In practice, however, it still confuses dialect and Standard German, and there are no reliable tests for spoken dialect. It is significantly more reliable for texts in the national languages.

Is Apertus as good as ChatGPT?

Not when it comes to complex tasks, math, programming, or source fidelity. The 70B variant performs well when answering questions in German and when translating. The real difference is transparency: You know what the model was trained on, and you can run it yourself.

Conclusion

Apertus isn’t a replacement for the major commercial models, but a genuine alternative when traceability, data sovereignty, and Switzerland’s languages matter. Version 1.5 adds long-context processing, Thinking Mode, and image understanding, and you can try it out for free today with Proton and Infomaniak. The best way to evaluate it: spend ten minutes testing it yourself with a real-life task from your daily routine.

In a nutshell:

  • Apertus is the fully open-source language model developed by ETH Zurich, EPFL, and CSCS, released on September 2, 2025.
  • Version 1.5, released on July 24, 2026, introduces 262,144 tokens, Context, Thinking Mode, and image understanding.
  • Try for free: Proton Lumo and Infomaniak Euria—no account required—at publicai.co/chat.
  • For businesses: Swisscom, Infomaniak, PHOENIQS, stepping stone, and Safe Swiss Cloud offer Apertus with data stored in Switzerland.
  • Locally, the 8B version requires about 6 GB, while the 70B version requires at least about 48 GB of storage.
  • Strong in openness, data sovereignty, and translation; weaker in mathematics, code, and fidelity to the source.

Are you already using Apertus? Tell us about it in the Friday Weekly—we're collecting real-world examples from the community.

As of October 6, 2026.

Sources: Apertus, official website · ETH Zurich on the launch of Apertus · CSCS on Apertus 1.5 · Proton on Apertus in Lumo · swissinfo: Apertus after one year · SRF on practical applications · Hugging Face, swiss-ai