Expanding Sovereign AI Choices

As part of its annual Birthday Week releases, Cloudflare has introduced new open-source artificial intelligence models to its Workers AI developer platform alongside hands-on security workshops for national cyber defense agencies. The company added EuroLLM, an open model covering all official languages of the European Union, and Apertus, Switzerland’s fully open multilingual AI model. The additions build on arguments Cloudflare made a year ago regarding national access to technology and digital sovereignty.

Writing on the official Cloudflare Blog, the company noted that discussions around AI sovereignty have intensified as geopolitical tensions grow and access to closed frontier models becomes increasingly restricted or geographically dependent. Cloudflare argued that rather than building restrictive technological walls, nations can enhance national resilience by adopting open-source models and decentralized architecture.

EuroLLM: Multilingual Coverage for Europe

The newly integrated EuroLLM project supports 35 languages, including all 24 official EU languages that have historically been underserved by proprietary open models. The model was developed by an international consortium including Instituto Superior Técnico, the University of Edinburgh, Instituto de Telecomunicações, Université Paris-Saclay, Unbabel, Sorbonne University, Naver Labs, and the University of Amsterdam. The project received backing from Horizon Europe, the European Research Council, and EuroHPC.

Trained on the MareNostrum 5 supercomputer, EuroLLM outperformed similar-sized models on EU multilingual benchmarks and machine translation tasks according to the consortium. Developers and public sector institutions can now request access to the model on Workers AI through Cloudflare's developer tools.

Apertus: Switzerland's Fully Open AI Architecture

Cloudflare also added Apertus, a fully open multilingual language model developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS) under the Swiss AI Initiative. Derived from the Latin word for 'open,' the project was built by public research institutions for public utility and aligns with European regulatory frameworks including the EU AI Act and GDPR.

Apertus was trained on CSCS's Alps supercomputer—utilizing more than 10,000 NVIDIA GH200 GPUs—across 15 trillion tokens in over 1,500 languages. Approximately 40 percent of its training dataset consists of non-English text. Its developers published its model architecture, weights, training methods, and data filtering procedures, designed specifically to respect training opt-outs, eradicate personal data, and avoid memorization. Performance evaluations show that Apertus significantly surpasses leading closed and open models on regional and rare languages, ranging from Romansh and Swiss German to low-resource languages across Africa and Asia.

Model-Agnostic Cyber Defenses for Infrastructure

Alongside model additions, Cloudflare announced hands-on training workshops aimed at helping government cyber agencies and critical infrastructure operators protect digital systems against AI-driven threats. The initial workshop is scheduled to launch in Singapore in October. The initiative stems from concerns that relying on a single frontier AI model for cyber defense leaves organizations vulnerable if access to that provider is suddenly restricted or restricted by geographical boundaries.

To address this vulnerability, Cloudflare's security team developed and open-sourced an orchestration harness layer. This system coordinates multiple AI models working in parallel to scan for vulnerabilities, verify security findings, and prioritize potential threats. By maintaining a model-agnostic architecture, defenders ensure their security posture remains operational even if access to an individual model provider is disconnected.

Real-World Applications Across the Asia-Pacific

The new European additions follow Cloudflare's earlier deployment of localized models from India, Japan, and Singapore. The company highlighted several regional applications created during buildathons and hackathons using these open models. In India, students at the Indian Institute of Technology Delhi worked alongside CyberPeace to build Form Mitra, a voice-guided assistant that helps citizens complete complex government forms in 22 Indian languages using the IndicTrans2 model.

In Singapore, developers created MedBridge, a clinical communication tool designed to guide patients through healthcare conversations across 14 languages and local dialects such as Hokkien and Cantonese, assisting foreign healthcare personnel. Meanwhile in Japan, developers built Anshin Concierge, an accessible single-tap platform that connects elderly residents and individuals with disabilities directly to appropriate Tokyo Metropolitan Government support desks based on plain-language descriptions of their issues.

Global Infrastructure and the Fight Against Lock-in

Cloudflare reiterated its commitment to preventing vendor lock-in across public sector and enterprise technology deployments. Company Chief Executive Officer Matthew Prince emphasized during the India AI Impact Summit 2026 in New Delhi that decentralized, affordable AI access forms an essential pillar of national resilience. The AI summit series is scheduled to move to Geneva in June 2027.

To support decentralized model deployment worldwide, Cloudflare has equipped GPUs across its global edge network, which operates in more than 335 cities across over 125 countries. More than 230 of these locations feature hardware acceleration dedicated specifically to AI inference, allowing governments and local developers to run regional open models close to end users without relying on a single corporate provider.

Sources

• Cloudflare BlogOne year later: Sovereign AI and the fight for choice