Open AI Models and Model-Agnostic Security Expand Choices for AI Deployment

AI sovereignty depends in part on whether organizations can change models without losing access to useful tools or security defenses. Cloudflare’s latest additions to Workers AI put that idea into practice: EuroLLM, a multilingual model developed by European research institutions, and Apertus, Switzerland’s fully open language model. The company is also starting hands-on workshops to help cybersecurity teams build defenses that can use different models.

The announcements follow a year in which access to some frontier models became more dependent on users’ location, while attackers began using advanced models in cyber attacks. Cloudflare argues that these shifts need not lead organizations to rely on a single provider or build systems around a single model. Its approach pairs open models with tools designed to work across providers, so a change in access does not automatically disable an application or a defense.

Two multilingual models join Workers AI

EuroLLM supports 35 languages, including all 24 official EU languages. Its developers say that many of those languages have been underserved by existing open models. A 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 developed the model with support from Horizon Europe, the European Research Council and EuroHPC.

EuroLLM was trained on the MareNostrum 5 supercomputer. The consortium reports that it outperforms models of similar size on EU multilingual benchmarks and machine translation. Those claims give developers a useful point of comparison, though the source does not provide specific scores or benchmark details. Access is available by request through Workers AI.

Apertus, whose name means “open” in Latin, was developed by ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS) through the Swiss AI Initiative. The model was trained on more than 15 trillion tokens spanning more than 1,500 languages; 40% of its training data is in languages other than English. Its developers published its architecture, weights, training data and methods, giving researchers more visibility into how the model was built.

The team trained Apertus on CSCS’s Alps supercomputer, which has more than 10,000 GH200 GPUs. The project says it took steps to respect training opt-outs, remove personal data and reduce memorization, with Swiss and European rules such as the EU AI Act and GDPR in mind. Developers report that Apertus significantly outperforms leading closed and open models on rare and regional languages, including Romansh, Swiss German and low-resource languages across Asia and Africa. As with EuroLLM, the source does not give scores for those comparisons. Access can be requested on Workers AI.

Local language support becomes practical

Cloudflare says hundreds of students, startups, small businesses and public servants across Asia-Pacific have built on models it added to Workers AI the previous year. Some projects began at buildathons run with local partners. They show how language coverage can help shape the design of a service, rather than simply add a translation feature.

• Form Mitra, India: Students at the Indian Institute of Technology Delhi built a voice-guided assistant to help people complete benefit forms written in dense English. The tool guides users in any of 22 Indian languages and uses AI4Bharat’s IndicTrans2. The project was developed at a buildathon with CyberPeace. Form Mitra is aimed at rural residents and people with low literacy or visual impairments.
• MedBridge, Singapore: Nurses caring for older patients may come from elsewhere in Southeast Asia and may not speak their patients’ local dialects. MedBridge guides health conversations in 14 languages, including Hokkien and Cantonese, to help reduce the chance that clinical information is lost in translation.
• Anshin Concierge, Japan: This hackathon prototype helps older residents, people with disabilities and others who find digital services difficult to navigate. Users describe a problem in everyday language, such as “my knees hurt” or “a strange screen appeared on my phone.” The tool points them to a relevant Tokyo Metropolitan Government support desk by phone or web page.

Security tools built to switch models

Cloudflare’s security team faced a related problem: a defense that relies on one model is only dependable while that model remains accessible. The team built a harness, an orchestration layer that coordinates multiple AI models working in parallel to search for vulnerabilities, verify findings and prioritize threats.

Cloudflare has shared its findings on combining frontier and open models, released the harness as open source, and described a layered defense design intended to make it harder for attackers using frontier models to find vulnerabilities. Since the harness can use open or closed models, organizations can change providers without switching off the entire workflow. The source does not specify the compute needs, operating costs or performance of the harness, so teams will need to assess those details in their own environments.

Cloudflare says discussions with cybersecurity teams led to requests for practical training on setting up the harness and adapting it to local requirements. Its new workshops will guide participants through building an AI security harness and layered defenses. The modules are designed to fit into existing AI skills and cyber resilience programs. The first workshop is scheduled for Singapore in October, during Singapore International Cyber Week.

More options as a measure of resilience

The wider case is about avoiding dependence on a single company or model. Cloudflare points to its work on open standards, free starting tiers and a network spanning more than 335 cities in 125-plus countries, with GPUs for AI inference in more than 230 countries. It presents those investments as part of a broader effort to make model choice usable in practice.

EuroLLM and Apertus add options for multilingual development, while the security harness applies the same principle to defense workflows. For teams evaluating them, the next questions are concrete: how the models perform on their own tasks, what deployment requires, and whether the harness fits their compute and operating constraints. The value of choice will depend on how well those tools work when put to use.