# What is Apertus, and why is it called a public AI model?

Date: 2025-09-03 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-apertus-swiss-public-ai-model/

**Answer:** Apertus is a large language model released on September 2, 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre, in sizes of 8 and 70 billion parameters. It is called public AI because public institutions built it for the public interest and published the weights, the training data recipes and the code, so anyone can inspect, reuse or run it.

Yesterday, September 2, 2025, three Swiss public institutions released Apertus, a large language model whose name is Latin for "open." A **large language model** is the kind of AI behind chat assistants: a program trained on huge amounts of text to predict what comes next. Plenty of companies now publish their models' final weights. Apertus goes further, and that is why it has drawn attention beyond Switzerland.

The release is a test of an idea more than a bid for the top of the leaderboard. Can public institutions, using public supercomputers, build something useful that others can examine from top to bottom?

## What was released

Apertus comes in two sizes, 8 billion and 70 billion **parameters**, the internal numbers a model adjusts during training. The press release from ETH Zurich suggests the smaller one for individual use. The model card on Hugging Face and the Public AI announcement give the licence as Apache 2.0, a permissive licence that allows commercial use.

The difference lies in what comes with the weights:

| Component | Typical open-weight release | Apertus |
|---|---|---|
| Final weights | Yes | Yes |
| Weights from partway through training | Rarely | Yes |
| Source code for training | Sometimes | Yes |
| Recipes to rebuild the training data | Rarely | Yes |
| Documentation of the process | Partial | Yes |

**Weights** are the trained numbers themselves. With weights alone, you can use a model but you cannot see how it got that way. With the data recipes and training code, a researcher can check what went in and try to reproduce the result. The developers also say Apertus was trained only on publicly available data, filtered to honour opt-out signals that websites publish in machine-readable form, even signals added afterward, and to remove personal data before training.

## Who built it, and where

Apertus comes from the Swiss AI Initiative, led by EPFL in Lausanne and ETH Zurich, working with the engineers and machines of the Swiss National Supercomputing Centre (CSCS). Imanol Schlag of ETH Zurich is the technical lead. Antoine Bosselut and Martin Jaggi of EPFL also speak for the project.

The funding is mainly public. The ETH press release cites over 10 million GPU hours on Alps, the CSCS supercomputer, along with support from the ETH Board and contributions from partners, above all Swisscom, the Swiss telecom company. A **GPU hour** is one graphics processor running for one hour, the usual unit for measuring AI training work.

Swisscom offers Apertus to its business customers on what it calls a sovereign Swiss platform. Outside Switzerland, access runs through the Public AI Inference Utility, a separate project that describes itself as the official international deployer. It says it has lined up more than 115,000 GPU hours across 20 clusters in more than five countries for September alone, with partners such as Amazon Web Services, Exoscale, AI Singapore and Australia's National Computational Infrastructure.

## How it works

Under the hood, Apertus is a standard text model that predicts the next piece of text, trained from scratch on 15 trillion **tokens** (small chunks of text). What stands out is the mix of data. About 40% of the training text is not in English, and the project includes languages that big models tend to neglect, such as Swiss German and Romansh. The sources do not quite agree on the count: ETH Zurich says more than 1,000 languages, and the Public AI announcement says more than 1,500.

The training ran in stages over web pages, code and math data, followed by a further round of adjustment for following instructions. The model card lists a context length of up to 65,536 tokens, roughly the amount of text the model can consider at one time.

One practical limit stands out. The model card says that no filter is available yet for removing personal information from the model's output, and it asks users to check back for one. Developers who build on Apertus should plan for that themselves, since a model can still produce personal details even when its training data was cleaned.

## Why it matters

A **fully open** model at this scale is rare, and the people behind it want it read as a blueprint. Martin Jaggi describes the aim as showing how "a trustworthy, sovereign, and inclusive AI model" can be developed. For a small country, the appeal is clear: expertise stays at home, and government or industry users do not have to rely wholly on a foreign supplier.

It also gives a concrete example to a growing movement for **public AI**, the idea that AI can be treated like public infrastructure. Joshua Tan of the Public AI Inference Utility compares it to "highways, water, or electricity."

The caution is that openness says nothing about quality. Independent results will take time. A 70-billion-parameter model is well below what the largest commercial labs are thought to train, and the developers themselves describe this as a first step, with regular updates planned. Other open efforts already exist too, such as BLOOM from the BigScience collaboration and OLMo from the Allen Institute for AI, so the fair comparison will be with those as well as with commercial products.

> **Try it:** Ask the same question in two languages you know, one of them a smaller language if you can, then ask a mainstream assistant the same thing. Note where Apertus is more natural, and where it simply knows less.

## How can you try Apertus yourself?

The easiest route is the chat interface of the Public AI Inference Utility, at [publicai.co](https://publicai.co/stories/apertus), which needs no installation. Swiss users can also try it through Swisscom and during the Swiss {ai} Weeks hackathons that are about to start.

If you want to run it, the [model page on Hugging Face](https://huggingface.co/swiss-ai/Apertus-70B-2509) has the weights and sample code. Start with the 8-billion version. As a rough rule, an unquantized model stored at 16 bits per parameter needs about 2 bytes per parameter, so about 16 GB for the small model and 140 GB for the large one, before any overhead. **Quantization**, which stores the numbers with less precision, can cut that sharply, at some cost in quality. The ETH press release adds that practical use requires additional pieces such as a server or a user interface, so a laptop test is a starting point more than a finished service.

A good first project is a small comparison: give it ten real questions from your own work and score the answers. That will tell you more than any announcement can.

## Sources

- Apertus: a fully open, transparent, multilingual language model: https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html
- With love, from Switzerland: https://publicai.co/stories/apertus
- swiss-ai/Apertus-70B-2509 model card: https://huggingface.co/swiss-ai/Apertus-70B-2509

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/what-is-apertus-swiss-public-ai-model/, no commercial use without permission, and keep the same licence.