The starting signal for AI occurred at the end of 2022. Since then, start-ups from all over the world have made a multitude of generative Artificial Intelligence models available to us. And the panorama of this new sector changes dramatically from one week to the next. Although, in fact, the AI race began about a decade ago, when ICT giants were looking for solutions to manage the huge amount of data generated on the Internet.
The protagonists are companies from the five continents, especially from the United States and China. Europe has been slow to enter this gold rush, which is much more than the fight for a market, since it has clear geostrategic implications. The vast majority of these models are proprietary, which has launched the debate about the lack of transparency. But there are open or semi-open models, which make new tools freely available to developers to refine, customize or create new AI models.
Semi-open AI systems
They are those that differ from the proprietary system because they are open source, that is, they allow access to the code. But on the other hand, they do not allow access to all the data, nor to the model training details.
Some of the limitations of these systems would be the difficulty of replicating the development process, the impossibility of preventing certain prejudices and discrimination, legal limitations on transparency issues in certain countries, as well as difficulties regarding collaboration between work teams.
Among others, we could mention Llama 3.1, from Meta, or GPT-OSS, from OpenAI.
Fully open AI systems
These systems do allow access to both the code and the training data and details. They are, therefore, much more versatile, reliable and transparent for developers. In this category we could mention EuroLLM, OLMo, SmolLM3,…, as well as Apertus, the model in question.
Apertus
In September 2025, the prestigious polytechnic universities of Lausanne and Zurich, EPFL and ETH respectively, as well as the Swiss Center for Scientific Computing, launched Apertus. It all started in December 2023. Since then, more than 800 researchers from 10 Swiss academic institutions have contributed to the project. The model runs on the Alps supercomputer, with more than 10,000 GPUs. These magnitudes make it one of the main exponents in terms of open Artificial Intelligence projects worldwide.
It is, therefore, a large-scale open model and its main users are currently Swiss SMEs and start-ups.
As characteristics to take into account with respect to other models, it is worth highlighting compliance at the data level, as well as multilingualism. Especially relevant is the effort that its creators have put into the diversity of languages. The model has been trained on 15 billion tokens across more than 1,800 different languages. These include Romansh and Swiss German, both minority languages. In fact, 40% of the model’s data is in languages other than English, which is a very high percentage, considering that Shakespeare’s language is widespread in the AI universe.
To access the model, there are two possibilities: downloading the Apertus LLM collection or open access to the chat version. The latter has been the subject of many criticisms and even ridicule, since, at first, its effectiveness was analyzed in comparison with that of other bots on the market. And the result was that other models gave more favorable results. Those responsible for the project were quick to clarify that the objective of Apertus is not to create a new GPT Chat. The main idea is to make the model available to developers. The chat version would be, at least for now, a sample button, proof that Apertus works.
On the other hand, by staying strictly within the RGPD and scrupulously respecting copyright, you will not provide information that involves violating the privacy or rights of third parties, which means important limitations on the possible responses.
Furthermore, if we compare again with Chat GPT, it reached 800 million users in April 2025, which implies that the model is fed by the millions of queries it receives, which obviously does not happen yet with Apertus.
The objective, therefore, is above all to be raw material and tool for new model makers. Apertus LLM collection allows you to choose between versions 8B and 70B, with 8,000 and 70,000 million parameters respectively, to choose from depending on the complexity of the projects.
Digital sovereignty
The appearance of this new model, which emphasizes transparency, ethics and solidity, acquires special relevance at a time when Europe is seriously considering its digital sovereignty.
Precisely at the end of October, the European proposal EDIC Digital Commons was made official, with the aim of developing open alternatives for AI, the cloud and cybersecurity. The member countries are France, Germany, the Netherlands and Italy; and with Luxembourg, Poland and Slovenia as observers.
Apertus is a large-scale open model and its main users are currently Swiss SMEs and start-ups.
This is not the first initiative of this type to be launched. Since 2021, the Gaia-X organization has existed, an association initially formed by eleven German entities and as many French entities and which has been growing exponentially since then.
Other initiatives, however, started with the same objective and ended as simple toasts to the sun. Some voices propose that the EU establish an agreement with Switzerland, since this country currently has a significant native digital infrastructure, in which Apertus is integrated.
In reference to this same topic, Byte TI magazine these days proposes the event Digital Sovereignty in Europe, which will bring together, on November 25 in Madrid, ICT experts around the problem of digital independence.
Be that as it may, everything would indicate that the newborn Apertus is destined to become, over time, an essential element in the panorama of global Artificial Intelligence, as well as the digital sovereignty of the Old Continent.
