As part of its traditional Birthday Week celebration of shipping digital presents to the internet, cloud infrastructure and security provider Cloudflare is introducing two prominent open-source European language models to its Workers AI platform. Developed by public universities and research institutions, EuroLLM and Apertus represent a major step forward in multilingual capability and regional digital sovereignty. EuroLLM covers all 24 official European Union languages alongside several others, while Switzerland’s Apertus is a fully open model trained across more than 1,500 languages. Access to both models is available for request starting today.
In addition to expanding model choice on its platform, the company is launching a series of hands-on cybersecurity workshops designed to help government cyber agencies and critical infrastructure operators build flexible AI defenses that function independently of any single model provider. The inaugural workshop is scheduled to take place in Singapore this October during Singapore International Cyber Week, addressing growing global concerns regarding AI security, geopolitical access restrictions, and infrastructure resilience.
These announcements build directly upon a foundational argument put forward a year ago when questions regarding artificial intelligence access and national sovereignty dominated discussions in government capitals worldwide. The core proposition then, as now, is choice: providing developers and organizations with the freedom to select the most appropriate tools for their specific tasks and the flexibility to switch between them as circumstances require.
Over the past year, however, those geopolitical and security conversations have intensified significantly. Malicious actors have increasingly leveraged frontier models to orchestrate sophisticated cyberattacks, while access to certain powerful models has become subject to geographic restrictions and shifting national policies. Concurrently, political calls to restrict open-source models have grown progressively louder. Taken together, these trends encourage a zero-sum perspective on artificial intelligence sovereignty—the assumption that every model controlled by another nation is one that cannot be relied upon domestically, leading to an instinctive urge to build protective walls and isolate local digital ecosystems.
Yet, industry observations from the past year suggest an alternative path forward. Nations such as India, Japan, and Singapore have leaned into open-source models, enabling developers and communities across the Asia-Pacific region to build practical tools tailored to rural citizens, elderly patients, and healthcare workers. Meanwhile, security engineering teams have successfully developed AI defense mechanisms that operate agnostically across multiple models, ensuring that losing access to a single proprietary provider does not compromise an organization’s overall security posture.
Maintaining an open, resilient internet has always depended on expanding options rather than restricting them. This philosophy underpins the development of open standards designed to prevent vendor lock-in, the creation of accessible developer tools, and the continuous expansion of a global network spanning more than 335 cities across over 125 countries, equipped with GPUs for AI inference in more than 230 of those locations. Within this framework, no country or enterprise should find itself entirely dependent on a single corporate entity for its artificial intelligence capabilities.
Two European Open Models on Workers AI
The integration of diverse, regional models aligns with broader international dialogues regarding decentralized technology. Earlier, leadership at major technology forums emphasized that decentralized, affordable access to artificial intelligence is fundamentally a matter of national resilience. The incorporation of models from India, Japan, and Singapore served as an initial validation of this approach, with upcoming international summits continuing the mission of fostering technology-driven prosperity and progress for all participating nations.
EuroLLM provides robust support for 35 languages, including all 24 official EU languages, many of which have historically been underserved by mainstream commercial open models. The model was developed with backing from Horizon Europe, the European Research Council, and EuroHPC, relying on a collaborative consortium that includes the Instituto Superior Técnico, the University of Edinburgh, Instituto de Telecomunicações, Université Paris-Saclay, Unbabel, Sorbonne University, Naver Labs, and the University of Amsterdam. Trained using the MareNostrum 5 supercomputer, the consortium reports that EuroLLM outperforms similarly sized models on European multilingual benchmarks and machine translation tasks. Developers can request access to the EuroLLM model on Workers AI through Cloudflare’s developer documentation.
Apertus, whose name translates from Latin as "open," represents Switzerland’s first large-scale, fully open, multilingual language model. Trained on more than 15 trillion tokens spanning over 1,500 languages—with approximately 40% of its training data dedicated to non-English languages—the project was brought to life by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre as a core initiative of the Swiss AI Initiative. Designed by public institutions for the public good, its complete architecture, weights, training data, and methodologies have been fully published.
Furthermore, Apertus was engineered to align closely with European regulatory standards, including the EU AI Act and GDPR, ensuring adherence to training data opt-outs, the removal of personal data, and the prevention of data memorization. Trained on the Alps supercomputer using more than 10,000 GH200 GPUs, the model’s developers indicate that it significantly outperforms leading closed and open alternatives when processing rare and regional languages, ranging from Romansh and Swiss German to various low-resource languages across Asia and Africa. Access to Apertus on Workers AI is likewise available by request.
What People Built Last Year
The national models introduced to Workers AI over the previous year have seen active implementation rather than remaining theoretical exercises. Hundreds of students, academic researchers, technology startups, small businesses, and public sector workers across the Asia-Pacific region have utilized these foundations to construct practical solutions, frequently collaborating during regional buildathons organized alongside local institutional partners. These real-world applications demonstrate the tangible value of localized, open-source artificial intelligence in addressing specific community needs.
AI Defenses That Don’t Depend on One Model
Governments increasingly seek to harness artificial intelligence to protect essential public services and critical national infrastructure. However, while advanced frontier models possess the scale required to identify complex software vulnerabilities, they can be utilized by adversaries just as effectively as by defenders. Furthermore, a security architecture built entirely upon a single proprietary model remains fragile, vulnerable to sudden restrictions in access or shifting geopolitical policies.
Internal security teams facing similar operational risks adopted a philosophy of acting as their own primary customers. Over the past year, these engineering groups collaborated across departments to construct AI-driven defense systems capable of functioning independently of any individual model provider. The resulting framework utilizes an orchestration harness—an administrative layer that coordinates multiple artificial intelligence models running in parallel to search for system vulnerabilities, verify analytical findings, and prioritize emerging threats.
By publishing insights regarding the combined use of frontier and open models, and open-sourcing the vulnerability orchestration harness, the organization has enabled external institutions to deploy the software alongside models of their own choosing. This layered architectural approach aims to block malicious actors equipped with frontier models from discovering vulnerabilities in the first place.
Because the security harness operates compatibly with both closed and open models, losing access to a specific commercial provider does not disrupt an organization’s defense mechanisms. Briefings with government representatives regarding this flexible architecture have consistently generated strong interest, leading directly to requests for structured, practical training programs.
To address this demand, the newly launched hands-on workshops are designed specifically for government cybersecurity agencies and critical infrastructure operators. Participants learn to construct their own artificial intelligence security harnesses and layered defense strategies, acquiring the practical knowledge necessary to adapt these systems to their specific organizational environments. Structured as modular, plug-and-play components, the training sessions fit seamlessly into established national digital skilling and cyber resilience initiatives, with the initial workshop commencing in Singapore during International Cyber Week.
Come Build With Us
The progress achieved over the past year has relied heavily on collaborative partnerships with organizations such as CyberPeace in India and Code for Japan, which have successfully translated open-source models into functional, real-world tools. Leaders of national artificial intelligence programs, cybersecurity agencies, and critical infrastructure operators seeking expanded technological autonomy and diverse operational options are encouraged to connect with policy teams to request access to EuroLLM and Apertus, or to explore the open-source security vulnerability harness.
As demonstrated over the past year of technological evolution, maintaining choice is not merely the pathway to digital sovereignty; it is equally essential for ensuring enduring artificial intelligence security.
Leave a Reply