Cloudflare Bolsters AI Sovereignty with European Open-Source Models
Cloudflare is integrating two new open-source AI models, EuroLLM and Apertus, into its Workers AI platform, enhancing user choice and promoting AI sovereignty, especially for government and critical infrastructure.
Cloudflare is expanding its Workers AI platform with the addition of two significant open-source AI models: EuroLLM and Apertus. This move aims to bolster AI sovereignty and provide users, particularly government cyber agencies and critical infrastructure operators, with greater choice and adaptability in their AI defense strategies.
EuroLLM, developed with support from European research institutions and funded in part by Horizon Europe, is designed to support 35 languages, including all 24 official EU languages. Its development consortium includes prominent universities and research bodies such as Instituto Superior Técnico, the University of Edinburgh, and Sorbonne University. Trained on the MareNostrum 5 supercomputer, EuroLLM reportedly outperforms similar-sized models on EU multilingual benchmarks and machine translation tasks, addressing a gap in AI model support for less common languages.
Complementing EuroLLM is Apertus, Switzerland's first large-scale, fully open, multilingual language model. Developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS), Apertus has been trained on over 15 trillion tokens across more than 1,500 languages, with a significant portion of its training data in non-English languages. Its architecture, weights, and training methodologies are publicly available, and it was developed with European regulations like the EU AI Act and GDPR in mind, focusing on data privacy and preventing memorization. Apertus has demonstrated superior performance on rare and regional languages, according to its developers.
These additions align with Cloudflare's year-old argument for AI sovereignty through choice, emphasizing the freedom to select and switch AI tools as needed. This approach contrasts with a growing trend towards nationalistic AI control and restrictions on open-source models, which some perceive as a necessary response to the increasing use of advanced AI in cyberattacks. Cloudflare argues that a focus on open-source models and decentralized access, as seen in initiatives from India, Japan, and Singapore, fosters national resilience rather than isolation.
The impact of accessible, open-source AI models is already evident. Cloudflare highlighted projects developed over the past year on models previously added to Workers AI. These include 'Form Mitra' in India, which uses AI to help rural citizens with benefit forms in their native languages; 'MedBridge' in Singapore, assisting healthcare workers and elderly patients with communication across various dialects; and 'Anshin Concierge' in Japan, which guides users to the appropriate public service based on their described needs.
Cloudflare's strategy extends to empowering organizations to build AI defenses that are not dependent on a single model. This flexibility is crucial for national security and critical infrastructure, ensuring that defenses remain robust even if access to specific AI models is disrupted or restricted. The company's global network, with GPUs for AI inference in over 230 cities, supports this vision of decentralized and accessible AI capabilities.
By integrating EuroLLM and Apertus, Cloudflare is not only expanding its AI offerings but also reinforcing its commitment to an open and diverse AI ecosystem. This move is expected to empower a wider range of organizations, particularly those in government and sensitive sectors, to leverage advanced AI securely and independently, fostering innovation while mitigating risks associated with proprietary or restricted AI models.
The availability of these European open-source models on Cloudflare's Workers AI platform signifies a step towards greater AI autonomy for nations and organizations, enabling them to build tailored AI solutions that respect local languages, regulations, and security requirements.