Apertus
A Swiss public large language model released in 8B and 70B sizes
Apertus is a public large language model put out on 2 September 2025 by the Swiss AI Initiative, a joint effort of EPFL, ETH Zurich and the Swiss National Supercomputing Centre. It ships in 8 billion and 70 billion parameter versions, covers more than 1800 languages, and is published under Apache 2.0 as an example of AI treated as a public good.
- Released 2 September 2025 in 8B and 70B parameter versions
- Trained on more than 1800 languages
- Weights published under the Apache 2.0 licence on Hugging Face
- Built by EPFL, ETH Zurich and the Swiss National Supercomputing Centre
- Described as meeting the EU Artificial Intelligence Act
What it is
Apertus is a publicly released large language model from Switzerland, made available on 2 September 2025. It is offered in two sizes, one with 8 billion parameters and a larger one with 70 billion, where a parameter is a learned weight inside the network. Training covered more than 1800 languages, so the system is aimed well beyond English. Its intended users are businesses and research groups around the world rather than casual chatbot users. The weights are published under the Apache 2.0 licence and can be downloaded from Hugging Face. The team behind it framed the release around multilinguality, openness and the ability for outsiders to inspect the system.
How it came about
The model came out of the Swiss AI Initiative, a partnership joining EPFL, ETH Zurich and the Swiss National Supercomputing Centre. It was announced on 2 September 2025 and put online the same day. The group positioned the work as AI built for the public rather than as a commercial product, and linked it to the idea of AI sovereignty, meaning a country or region keeping control of the technology it depends on. The first release was described as a starting point that would need adaptation for particular tasks.
Inside the model
Apertus is distributed in two parameter counts, 8 billion and 70 billion. Its training corpus spans more than 1800 languages, which is unusually broad and points to a deliberate focus on multilinguality. The build was guided by an intention to stay within European copyright rules. At launch the creators named three priorities: support for many languages, transparency about how the system works, and auditability so that outside parties can examine it. Those goals were set out in contrast to the closed commercial frontier systems. Fine-tuning and alignment details were not given in the release notes available here.
What it can do
Coverage of over 1800 languages is the standout capability. The model was built first for business and research settings worldwide. Reviewers noted that the opening version sits well behind the strongest frontier systems in raw capability. For many real tasks it has to be adapted before it is useful. Chatbots are treated as a secondary application rather than the main one, so conversational use is not the design centre.
How to use it
Both sizes are downloadable from Hugging Face under the Apache 2.0 licence, which permits reuse and redistribution. The project also has a public website at apertus-ai.org. Because the weights are openly published, anyone can run the model themselves rather than relying on a hosted service.
Safety and testing
The release stressed transparency and auditability as ways for outsiders to check the system. The model was built with the aim of following European copyright law. Beyond those points, no separate evaluation or red-teaming programme is described here.
Reception and debate
Coverage outside Switzerland was mostly positive. The same reporting noted that the first version trails frontier models by a wide margin and that most uses need further adaptation. Chatbot behaviour was seen as a side application, not the main draw.
Why it matters
Apertus is cited as one of the early cases of a language model treated as a public good, tied to the AI sovereignty argument. It is also described as the first large model to meet the requirements of the European Union's Artificial Intelligence Act. Its emphasis on multilinguality, transparency and auditability offers a different set of priorities from the closed commercial systems.
Questions people ask about Apertus
Who is behind Apertus?
It comes from the Swiss AI Initiative, a partnership of EPFL, ETH Zurich and the Swiss National Supercomputing Centre. The model was released on 2 September 2025 and is run under the Swiss AI name.
How big is it?
Two versions exist: one with 8 billion parameters and one with 70 billion. Parameters are the learned weights that make up the network, and the larger figure is the top end of the range offered.
Can I download and use it?
Yes. Both sizes are on Hugging Face under the Apache 2.0 licence, which allows reuse and redistribution. There is also a project site at apertus-ai.org.
How does it compare with frontier models?
The first release is reported to sit well behind the strongest frontier systems. Most practical uses need adaptation first, and chatbots are a secondary application rather than the main target.
What makes it different?
Training covered over 1800 languages, and the team stressed multilinguality, transparency and auditability. It is presented as AI as a public good and is described as meeting the EU Artificial Intelligence Act.
Good reading and sources
- Official websiteOfficial site
- Switzerland releases an open-weight AI modelThe Verge, Article
- The Race to Build the DeepSeek of Europe Is OnWired, Article
- swiss-ai (Swiss AI Initiative)Hugging Face, Article
The facts on this page come from the Wikipedia article "Apertus (LLM)", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.