Switzerland's Apertus AI model surpasses 4 million downloads in one year

Summary

Switzerland's open-source large language model (LLM), Apertus, surpassed 4 million downloads on Hugging Face less than a year after its release in September 2025. The model has found concrete applications in multilingual government translation and medical research. However, it lags significantly behind US commercial models in coding, mathematics, and complex automation.

Why it matters

Apertus is one of the rare large-scale initiatives to build a truly open, multilingual, and data-protection-compliant foundation model outside the US–China duopoly. Forty percent of its training data consists of non-English languages, covering over 1,800 languages. Its adoption by the Canton of Ticino, AI Singapore's SEA-LION project, and EPFL's MeditronFO project demonstrates that sovereignty and openness can drive genuine institutional demand. The 2027 roadmap for version 2.0 indicates that Swiss public research institutions are committed to long-term AI infrastructure rather than a one-off release.

Highlights

• Over 4 million downloads and deployment in 70+ foreign countries in the first year, including Singapore's SEA-LION adaptation
• ~40% non-English training data, compared to the ~10% typical in commercial LLMs; covers 1,800+ languages, including Romansh
• Lags behind US models in programming, mathematics, and complex automated tasks; version 2.0 is not expected until 2027

Critical take

The article relies too heavily on download numbers and institutional endorsements. In reality, the model was deemed 'unusable' for basic questions just weeks after its release. The 4 million figure on Hugging Face conflates research curiosity with production-level adoption. The delay of a competitive version 2.0 to 2027 means that US and Chinese labs will likely release multiple versions during this two-year gap.