Special Edition: Apertus – The Swiss AI

September 5, 2025

Executive Summary

Switzerland has entered the global AI race with Apertus, the country’s first large language model developed through a collaboration between ETH Zürich, EPFL, and the national supercomputing center CSCS in Lugano. Launched in September 2025, Apertus represents more than a technological achievement – it is an attempt at establishing digital sovereignty and regulatory compliance in an increasingly fragmented AI landscape.

Unlike the “open-weight” models we discussed in our Mistral edition, Apertus adopts a “fully open” approach, publishing training data, code, and even intermediate checkpoints. This transparency-first philosophy, combined with its stated goal of compliance-by-design architecture, positions it as a potentially attractive option for regulated financial institutions, though current performance limitations require realistic expectations.

Initial hands-on testing reveals a functional but still developing system that excels in multilingual capabilities – including Swiss German – while still having ground to cover compared to commercial alternatives in complex reasoning tasks.

Switzerland’s Strategic AI Move

The Apertus project emerged from the Swiss AI Initiative, launched in December 2023 as a collaboration between the country’s leading technical institutions. The CHF 20 million investment from the ETH Board, combined with over 10 million GPU hours on the “Alps” supercomputer in Lugano, represents Switzerland’s most significant commitment to sovereign AI development.

Key players and infrastructure:

Academic consortium: ETH Zürich, EPFL, and CSCS bringing together more than 800 researchers
Strategic partner: Swisscom provides the commercial platform for business deployment
Technical foundation: Training on Alps supercomputer using up to 4,096 of the available 10,000 Nvidia GH200 GPUs
Data scale: 15 trillion tokens across more than 1,000 languages

The timing is deliberate. As we observed in our July insights on the AI for Good Summit, the global AI landscape is consolidating around US and Chinese ecosystems. Apertus represents Europe’s attempt to establish a “third way” – one that prioritizes transparency, compliance, and regional values over pure performance metrics.

“Fully Open” – Beyond Mistral’s Approach

Where Mistral AI pioneered the “open-weight” model by releasing trained parameters, Apertus goes further with complete transparency. The project publishes:

Complete training datasets and data preparation scripts
Full source code and training recipes
Intermediate checkpoints showing learning progression
Evaluation tools and benchmarking results

This radical openness contrasts sharply with both commercial models and previous European alternatives. While Abu Dhabi’s TII Falcon models offer open-source capabilities, and China’s telecom operators provide enterprise-ready solutions, Apertus enables full reproducibility and auditing – critical features for financial institutions operating under strict regulatory oversight.

The Apache 2.0 license removes commercial restrictions, allowing unrestricted business use while maintaining the transparency that regulators increasingly demand.

First Hands-On Test: Promise and Reality

We conducted initial practical tests of Apertus across both online platforms and local infrastructure to assess real-world viability for wealth management applications.

Performance findings:

Functionality: The model operates reliably for basic tasks including text analysis and translation.
Speed: The first “downloadable” version appears to us to be relatively slow on local infrastructure. We expect this to improve as other technical variants and configurations become available.
Language capabilities: Strong performance in German and notably functional Swiss German – potentially valuable for client-facing applications in the Swiss market
Technical deployment: Successfully ran on standard enterprise hardware (8B parameter version runs on a modern laptop), though resource requirements are substantial for the 70B parameter version

Business-relevant observations: The multilingual strength extends beyond marketing claims. Swiss German processing opens interesting possibilities for automated client communication and document analysis in the domestic market – a capability that US-centric models may sometimes handle suboptimally. However, for sophisticated investment research or complex analytical tasks, performance currently lags behind established alternatives from larger US tech players.

Strategic Implications for Swiss Wealth Managers

Apertus could present a unique value proposition that wealth managers should evaluate carefully, particularly in light of regulatory developments.

Immediate advantages:

Data sovereignty: Processing remains within Swiss/EU jurisdiction when using Swiss/EU server or own local infrastructure
Regulatory compliance: Designed for adherence to GDPR and Swiss data protection laws
Transparency: Full auditability addresses increasing regulatory scrutiny of AI systems
Local optimization: Swiss German capabilities for domestic client interactions

Current limitations:

Performance gap: Complex analytical tasks better served by commercial alternatives
Resource requirements: Significant infrastructure needs for self-hosting
Limited specialization: No finance-specific fine-tuning yet available

Practical application scenarios: Apertus appears best suited for compliance-sensitive applications where transparency trumps peak performance: regulatory reporting, client communication in local languages, and document processing where data residency is critical. For investment research, portfolio analysis, or sophisticated client advisory support, established commercial models remain more effective for the moment.

EU AI Act Context: A Compliance-Ready Foundation

Apertus aims to be one of the most regulation-ready large language models available to European financial institutions.

Regulatory positioning: Where our August newsletter highlighted the compliance challenges facing firms using ChatGPT variants, Apertus eliminates many concerns by design. The stated complete transparency of training data, the respect for opt-out requests, and the documentation aimed at addressing EU AI Act requirements could create a lower-risk option for regulated use cases.

We will explain the EU AI Act in one of our upcoming webinars.

Looking Ahead

Apertus represents an important strategic development rather than an immediate game-changer. Its real-world impact will depend on how Swiss researchers, businesses, and institutions leverage its unique capabilities – the potential is clearly there.

Switzerland has successfully established a sovereign AI capability that prioritizes transparency and compliance over raw performance. For wealth managers, this creates a new option in the vendor landscape – one particularly suited for applications where regulatory safety outweighs peak capabilities.

The project’s roadmap toward domain-specific models (mentioned for legal, healthcare, and climate applications) suggests future finance-specific versions that could better serve our industry’s needs. Until then, Apertus serves as a valuable compliance-focused alternative for specific use cases, while the broader AI toolkit will likely continue relying on commercial solutions for performance-critical applications.

The broader lesson for our industry: the AI landscape is evolving beyond a simple US-China duopoly. European alternatives like Mistral and now Apertus offer differentiated approaches that may prove valuable as regulatory requirements tighten and digital sovereignty concerns grow.

Sources:

About the Author: Dr. Andreas K. Janoschek specializes in AI applications for Asset & Wealth Management. Based in Geneva, he helps industry professionals navigate the intersection of finance and technology.

This newsletter aims to inform and does not constitute investment or legal advice. Always consult with qualified professionals for specific circumstances.