The development of open AI models is gaining momentum in Europe. Even so, the continent still lags significantly behind China and the United States. This is shown by a new white paper from Agora Digitale Transformation. The study examined 1,037 foundation models for text generation with at least seven billion parameters that were published on the Hugging Face platform between January 2023 and September 2026. The new Mistral Large 4 model is not yet included.
Of the models analysed, 559 came from China, 369 from the United States and just 109 from Europe. European models therefore account for around 10%. However, their number is growing particularly quickly: by September 2026, almost 70% more European models had been released than in the whole of the previous year. In absolute terms, however, the gap remains substantial: around 40 new European models compare with approximately 140 from China.
According to the study, Europe’s strength lies primarily in medium-sized AI models. Just over one third of European models have between 30 and 130 billion parameters. Such models are particularly well suited to public administration and SMEs and represent a “third option” alongside US and Chinese AI offerings, according to study author Yannick Kalff. To strengthen this position, he recommends measures including public-private partnerships, research funding and European distribution platforms.
Kalff also argues that greater transparency around training data and source code could become a European quality marker. So far, however, only 45% of European models are permissively licensed, meaning that third parties are allowed to use, modify, copy and commercially exploit them.
By Maria Kleiner
Source: Tagesspiegel Background, 4 November 2026, by the editorial team
