Switzerland launches the language model Apertus | Press review n°22

I selected important news on artificial intelligence during the week of September 1st to 7, 2025. Here's my commentary.

Sep 08, 2025

Illustrative photo by Kelly Sikkema on Unsplash

Welcome to the twenty-second press review of Artificial reality. This week I focused on a new artificial intelligence model from Switzerland, on a controversial data analysis company, and on the impact of AI on work. Have a good read!

📰 Read

Switzerland launches the language model Apertus

The Federal Polytechnic School of Lausanne (EPFL) announced on Tuesday a new language model called Apertus, developed in partnership with the Federal Institute of Technology Zurich (ETH Zurich) and the Swiss National Supercomputing Centre (CSCS), Le Temps reports.

Language models are an artificial intelligence technology on which the conversational agents such as ChatGPT are based. The chatbot of OpenAI, for example, uses different language models like GPT-5 or GPT-4o, each having its own characteristics. These models are trained on vast datasets, including texts, to learn how to respond to users’ questions in natural language.

The name of the Swiss language model, Apertus, means “open” in Latin. It reflects the fact that this AI model is open‑source, meaning its parameters and code are publicly available and any company, public administration, or individual can use it freely and at no cost. Moreover, it was trained only on openly licensed data that are not protected by copyright.

Apertus: a fully open, transparent, multilingual language model (EPFL)

An alternative to tech giants

Antoine Bosselut, director of the Natural Language Processing Laboratory at EPFL and co‑leader of the Swiss AI Initiative, an organization of the federal institutes of technology and the Swiss Confederation, says it is important for Switzerland to have its own artificial‑intelligence model.

“In certain domains, [AI] surpasses human performance, which shows its power. But beware: these services and language models are built by a handful of tech giants who keep their designs and training data secret. That is a huge problem. We cannot accept that the foundations of such powerful AI are in the hands of a few multinationals,” he says to Le Temps. Apertus therefore aims to be an alternative to the commercial models already on the market.

EPFL, ETH Zurich, and CSCS have committed the resources needed to fulfill this ambition. Around 150 engineers and professors developed Apertus, leveraging the power of the Alps supercomputer in Lugano. This machine, one of the world’s most powerful, is equipped with 10,000 state‑of‑the‑art graphics processors and cost roughly 100 million Swiss francs (around $125 million). Overall, the price tag for this language model is between 5 million and 10 million Swiss francs ($6-12 million), covering the hardware, electricity, and salaries.

Apertus, having been trained on 1,000 languages, is intended for an international use. It is released in two versions, one with 8 billion and another with 70 billion parameters, similar to Meta’s Llama 3 model that was launched in April 2024.

“The results are not extraordinary”

The collective Public AI, which supports public artificial intelligence projects, quickly released a chatbot based on Apertus, at https://publicai.co. It is free to use.

Le Temps journalist Anouch Seydtaghia tested it. In an article published the day after Apertus’ launch, he regrets that “the results are not extraordinary.” He gives several examples:

“The chatbot based on Apertus claims, for instance, that the EPFL is located in the municipality of Pully (when it is actually west of Lausanne). It also says that Guy Parmelin was born in Bursins, ‘a Geneva municipality’ (while Bursins is in the canton of Vaud) and that ‘his election as president is remarkable because he becomes the first President of the Swiss Confederation from the canton of Vaud in over a century’ (which is false, because there was also Jean‑Pascal Delamuraz).”

“These errors don’t worry me”

Other criticisms of Apertus have appeared online, notably on LinkedIn. One comment reads: “Sorry, this LLM is a disgrace… we need a lot more rigor, especially from our EPF… This LLM is the worst ever tested at launch… hallucinations are frequent,” wrote Xavier Comtesse, president of the think tank Manufacture Thinking.

Interviewed in an article from Le Temps published on Friday, Steph Cruchon, the director of Design Sprint, an innovation workshop based in Lausanne, says that “in innovation, to achieve anything ‘good’ you have to go through several phases: prototyping, testing, a minimal viable version, more testing, further versions… that’s a completely normal process. What isn’t normal is releasing a product this early. After just two queries you notice a glaring lack of training time or accessible data to train the model. Why wasn’t it kept internally longer?”

Antoine Bosselut remains optimistic: “These errors don’t worry me, I understand there are very high expectations for Apertus,” he tells Le Temps. “We will continuously improve our model and we are listening to every feedback,” he adds.

An ethical and political milestone

Nicolas Naully, a specialist in integrating generative‑AI services, believes Apertus offers qualities beyond raw technical performance:

“Sure, if you bluntly compare Apertus to ChatGPT or Claude, the technical gap is obvious. But what truly matters is the quality of output on validated, contextualized datasets. There’s also the application of responsible AI principles from the ground up: respect of intellectual property, transparency about data, and integrated linguistic and cultural diversity. Finally, providing an open‑source foundation for Swiss and European use, without dependence on American giants, sets an ethical and political milestone that many critics overlook.”

Le Temps notes that many observers praise Switzerland’s willingness to propose an alternative to the services of big AI companies. This new model rests on solid ethical foundations but still needs refinement, according to the general consensus. As Gregory Mermoud, associate professor at the HES‑SO Valais School of Computer Science, puts it, “the real test will be the Apertus team’s ability to keep iterating and improving, with budgets that remain alive after a launch that could be described as chaotic.”

Watch (in French): Des chercheurs suisses lancent le premier modèle IA transparent et ouvert à tous (RTS) and Des chercheurs suisses lancent Apertus, un modèle IA de langage ouvert: interview de Martin Rajman (RTS)

Seven important news this week

OpenAI Plans to Add Safeguards to ChatGPT for Teens and Others in Distress (The New York Times)
FTC Prepares to Question AI Companies Over Impact on Children (The Wall Street Journal)
Is It Safe to Upload Your Photos to ChatGPT? (The Wall Street Journal)
Tech leaders take turns flattering Trump at White House dinner (The Verge)
Congress Pushes DHS for Details on ICE’s New Facial Recognition App (404 Media)
ICE to Gain Access to Spyware After Biden Order Dropped (Bloomberg)
Army’s Contract With Startup to Give Soldiers Battlefield AI (The Wall Street Journal)

🎥 Watch

AI benefits Palantir

In this video published on Friday, Vanessa Wingårdh talks about Palantir, a controversial multinational which proposes data analysis services to public administrations, intelligence agencies, armies, and companies, with the help of artificial intelligence.

Palantir, whose European headquarters are in Switzerland, is currently experiencing explosive growth, with its stock up more than 600% from a year ago. In the United States, its first global market, the company is suspected of building a massive database on the population, raising concerns regarding mass surveillance and privacy.

In her video, Vanessa Wingårdh provides an analysis of Palantir, from its funding to its current activities worldwide.

How Your Tax Dollars Built Palantir's Global Surveillance Empire

Seven important videos this week

Trump hosts top tech CEOs, not including Elon Musk, at White House (PBS)
ICE can hack your phone without you knowing (Taylor Lorenz)
Real Life "Skynet" (Upper Echelon)
We Investigated Tesla’s Autopilot. It’s Scarier Than You Think (More Perfect Union)
This $7.5 Billion Startup Says AI Surveillance Is The Future Of Fighting Crime (Forbes)
Weapons company founded by 21-year-old lands a U.S. Army contract (NBC News)
Will Humanoid Robots Take Our Jobs? (Interesting Engineering)

🔈 Listen

Will AI kill your job?

In the new episode of Tech Won’t Save Us, journalists Paris Marx and Brian Merchant talk about the disappointing release of GPT-5 and whether the AI bubble is about to burst. They also address the impact of AI on jobs.

Listen to Tech Won’t Save Us on Patreon