One year later: Sovereign AI and the fight for choice

Two European open models, EuroLLM and Apertus, are being made available on Workers AI. EuroLLM supports 35 languages, including all 24 official EU languages, and was developed with contributions from various European institutions and trained on the MareNostrum 5 supercomputer. Apertus is Switzerland's first large-scale, fully open, multilingual language model, trained on over 15 trillion tokens across more than 1,500 languages using resources from ETH Zurich, EPFL, and CSCS. These models are being offered to the public via Workers AI access requests.
The initiative also includes hands-on workshops for government cyber agencies and critical infrastructure operators focused on building AI defenses that operate across any model. The context suggests a shift in the discourse around AI access, moving from concerns about access control to focusing on AI sovereignty through choice and building resilient defenses. This approach is framed by the observation that restricting access to frontier models creates an imbalance where security depends on a single provider, prompting a push for decentralized, multi-model defense strategies.
Furthermore, the initiative highlights practical applications where these models have enabled local solutions, such as Form Mitra in India for government forms, MedBridge in Singapore for healthcare translation, and Anshin Concierge in Japan for public service access. The underlying philosophy emphasizes open standards to prevent vendor lock-in, connecting AI access with broader goals of national resilience and building a more inclusive internet.

* EuroLLM covers all 24 official EU languages.
* Apertus is Switzerland's first large-scale, fully open, multilingual language model trained on over 15 trillion tokens across more than 1,500 languages.
* EuroLLM was developed with support from a consortium including Instituto Superior Técnico, the University of Edinburgh, and others, and trained on the MareNostrum 5 supercomputer.
* Apertus was developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS).
* Apertus respects EU rules like the EU AI Act and GDPR regarding training data.
* The initiative includes hands-on workshops for government cyber agencies and critical infrastructure operators to build multi-model AI defenses.
* Form Mitra in India is a benefit form assistant using an AI4Bharat model for voice guidance in 22 Indian languages.
* MedBridge in Singapore guides patients through health conversations in 14 languages, including Hokkien and Cantonese.
* Anshin Concierge in Japan connects elderly residents to public services based on user description.
* The initiative promotes open standards to prevent vendor lock-in and provides access via Workers AI.

The narrative constructs a powerful argument that redefines AI sovereignty from a zero-sum control battle into a security and agency issue. The transition from focusing on centralized model control to advocating for model choice positions the decentralized nature of open models not merely as an alternative, but as a necessary precondition for national resilience. The implied mechanism is that dependence on a single entity creates systemic vulnerability, directly linking access limitations to national security risks, which provides a strong impetus for change.
The juxtaposition of the high-level technical achievement (training massive multilingual models) with tangible, localized societal benefits (assisting rural citizens and healthcare workers through dialect-specific tools) serves to bridge the gap between abstract technology and human dignity. This move from pure technological capability to applied public good is a sophisticated strategy for building public trust.
A critical tension exists in how the call for open access interacts with existing security imperatives. The argument suggests that distributing models mitigates external attack vectors (by allowing defenses built on any model), yet the immediate reaction often centers on operational security. This sets up an implicit dilemma: whether optimal sovereignty is achieved through maximal decentralization, or through resilient, layered systems that can incorporate whatever tools are available. The focus shifts from *who controls the tool* to *how we architect defense regardless of the tool's origin*.
The patterns observed suggest a deliberate pivot away from purely proprietary control toward community-driven infrastructure and shared security protocols. This mirrors historical movements where access is granted not through concession, but through demonstrable utility and shared ownership, suggesting a move toward an entitlement-based model for foundational AI infrastructure rather than a simple licensing arrangement.
Bridge Questions: If the goal of sovereignty is resilience, what specific governance structures are necessary to manage the inevitable conflict between open development and national security requirements? How can the emphasis on building defenses with *any* model be operationalized into concrete, measurable national cyber defense standards? What happens when the incentive for open research conflicts with the need for centralized, coordinated threat response during a crisis?