EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus, a fully open model intended as a foundation for research and applications. Versions with eight billion and 70 billion parameters were made available under a permissive licence.
Architecture, model weights, training information, documentation and intermediate checkpoints were opened for inspection. Developers could download the model or access it through Swisscom and the Public AI Inference Utility for experimentation and hosted use.
Training used 15 trillion tokens in more than 1,000 languages, with about 40 percent of the data in languages other than English. Swiss German and Romansh were included among languages frequently underrepresented in large models.
The project filtered public information to remove personal data and honour website opt-outs, while following Swiss data protection, copyright rules and EU AI Act transparency requirements. Future work targeted efficiency and domain tools for law, health, climate and education.
The development should be read in its original publication context. It described the evidence, forecasts and vendor claims available at that time; later outcomes should not be projected backwards into promises that had not yet been tested.
A useful assessment separates demonstrated functions from proposed scale. It compares the new approach with a realistic baseline, including manual work, waiting time, existing infrastructure and the cost of correcting errors.
Long-term value requires more than launch performance. Data, models, organisations and external rules change, so version records, periodic checks and a named owner are needed to detect drift and decide whether the original case still holds.
Distribution also matters. An average gain can conceal predictable weakness for a language, site, workload or user group. Breaking results down and inviting feedback from people closest to the process makes expansion decisions more defensible.
Before scaling, leaders should map the complete delivery chain. Data sources, interfaces, cloud services, suppliers and human review can each become the limiting factor, even when the most visible model performs well. That map also exposes concentration, security and recovery requirements that a product demonstration rarely shows.
Financial analysis should follow outcomes rather than activity. More generated content, processed alerts or automated steps is not automatically more value. Completed tasks, avoided incidents, lower correction effort and improved service quality give decision-makers a stronger basis for continuing investment.
Reversibility is equally important in a fast-moving market. Portable records, documented interfaces and clear stop criteria let an organisation change provider, narrow a deployment or return a decision to people when performance, pricing or policy changes. That flexibility is practical resilience, not a lack of commitment.
Full Openness Extended Beyond Model Weights
Apertus exposed data documentation and checkpoints as well as the final parameters. That allowed researchers to study training choices and reproduce or adapt more of the development process.
The practical importance was that openness increased verifiability but also required clear provenance, secure distribution and documentation that non-specialists could understand. A credible evaluation would connect the claim to representative outcomes, record exceptions and human intervention, and identify who could act when conditions changed. It should compare different workloads or user groups and document the resources required to sustain the result, preventing one headline metric from substituting for operational evidence. Teams would also need a baseline, a review schedule and a defined threshold for pausing or changing the system. Those controls would make progress measurable while preserving room to respond when costs, risks or user needs diverged from the original assumptions.
Multilingual Training Broadened Public Value
More than 1,000 languages and a large non-English share aimed to serve communities neglected by models dominated by English data. Local languages also supported Switzerland’s public-interest and sovereignty goals.
The practical importance was that coverage statistics needed evaluation of fluency, factual reliability and cultural context in each language rather than assuming token presence guaranteed quality. A credible evaluation would connect the claim to representative outcomes, record exceptions and human intervention, and identify who could act when conditions changed. It should compare different workloads or user groups and document the resources required to sustain the result, preventing one headline metric from substituting for operational evidence. Teams would also need a baseline, a review schedule and a defined threshold for pausing or changing the system. Those controls would make progress measurable while preserving room to respond when costs, risks or user needs diverged from the original assumptions.
Hosted Access Complemented Downloads
Experienced teams could run the models themselves, while Swisscom and a public inference utility lowered infrastructure barriers. Hackathons offered another route to practical testing.
The practical importance was that public access remained meaningful only if pricing, capacity and interfaces allowed schools, researchers and smaller organisations to participate. A credible evaluation would connect the claim to representative outcomes, record exceptions and human intervention, and identify who could act when conditions changed. It should compare different workloads or user groups and document the resources required to sustain the result, preventing one headline metric from substituting for operational evidence. Teams would also need a baseline, a review schedule and a defined threshold for pausing or changing the system. Those controls would make progress measurable while preserving room to respond when costs, risks or user needs diverged from the original assumptions.
Compliance Was Treated as a Design Principle
Filtering personal information, respecting opt-outs and documenting training aligned openness with legal responsibility. The project argued that transparency and performance did not have to be opposing choices.
The practical importance was that independent audits and future dataset updates would test whether those commitments survived continued model expansion. A credible evaluation would connect the claim to representative outcomes, record exceptions and human intervention, and identify who could act when conditions changed. It should compare different workloads or user groups and document the resources required to sustain the result, preventing one headline metric from substituting for operational evidence. Teams would also need a baseline, a review schedule and a defined threshold for pausing or changing the system. Those controls would make progress measurable while preserving room to respond when costs, risks or user needs diverged from the original assumptions.
Across the source report, technical capability was only one part of the system. Skills, access, integration, governance, cost and affected people could each determine whether the promised value survived outside a demonstration.
The durable lesson is to preserve traceability and choice. Organisations should document assumptions, keep consequential outputs reviewable, monitor downstream effects and retain a realistic exit route. Those controls allow ambitious AI work to improve without making early momentum an irreversible commitment.
