EU Bets on EUROPA Consortium to Build Its Own Frontier Model: 400B Parameters, 24 Languages, Open Source
The US and China have been in an AI arms race for three years. Europe has finally decided to build its own.
Three key facts
On June 19, the European Commission announced that the EUROPA consortium, led by Italian company Domyn, has won the Frontier AI Grand Challenge. The competition launched in February 2026, requiring participants to propose a model exceeding 400 billion parameters with native coverage of all 24 official EU languages. EUROPA will receive 2.5% of the total EuroHPC computing capacity for one year on one or more AI-optimized supercomputers. The final model will be released as open source.
The technical approach centers on Mixture-of-Experts (MoE) architecture. According to the Commission's announcement, EUROPA plans to use efficient, modular MoE architectures to achieve frontier-level performance with limited compute. This is not surprising — Mistral and DeepSeek have already proven that MoE is the best path for building large models on a budget. But 400 billion parameters is not that large for MoE; the real challenge is hitting frontier performance simultaneously across 24 languages, many of which are low-resource languages with extremely scarce high-quality training data.
This is the latest step in the EU's digital sovereignty strategy. For years, the EU has played the role of AI regulator — producing the world's first AI Act — while remaining heavily dependent on American companies for actual model development. EUROPA's goal is clear: train a frontier open-source model on European infrastructure, with European data and talent, free from US corporate control, for use by European businesses, researchers, and public institutions.
WangDou's Take
The EU's problem has never been money or talent — it's speed. Launch the challenge in February, pick a winner in June, then another year of training — by the time this model ships, GPT-6 and Claude 5 might have already gone through two iteration cycles. And 2.5% of EuroHPC's total compute sounds generous, but NVIDIA is reportedly supplying a 6,000-card Blackwell cluster, while Meta trained Llama 3 on 16,000 H100s. Trying to cover 24 languages at 400 billion parameters with that kind of compute is like entering an F1 race in a minivan. That said, Mistral built Mixtral with far less resources than OpenAI, and DeepSeek trained R1 on just 2,000 cards. When compute falls short, MoE fills the gap — if EUROPA actually delivers a 24-language open-source powerhouse, native support for Estonian and Maltese alone would be enough to make GPT-5 sweat.