Cloudflare brings publicly built EuroLLM and Switzerland's Apertus to Workers AI

Apertus's compliance work happened at training time. According to Cloudflare's post, designing it around the EU AI Act and GDPR meant respecting training opt-outs, removing personal data and preventing memorization ^[5]. Those are properties of the weights; where a prompt is processed depends on the host. Workers AI has inference GPUs in more than 230 of the more than 335 cities on Cloudflare's network ^[10]. The post does not say whether requests to either model are kept on European GPUs.

The stronger sovereignty case is the exit. Cloudflare argues against dependence on any single provider, itself included. "We don't think any country should have to depend on one company for its AI. That includes us," the company wrote ^[11]. For Apertus that sentence can be tested, because its architecture, weights, training data and methods are all published ^[4]. A team that builds against the Workers AI endpoint can later run the same model on its own hardware. The post calls EuroLLM open as well ^[16]. A team counting on the same exit for EuroLLM should confirm which of its artefacts are downloadable before writing code against it.

The Apertus build is careful work. Forty percent of more than 15 trillion training tokens ^[3] is more than 6 trillion tokens outside English ^[1]. Because the corpus is published alongside the weights ^[4], an outside auditor can check the opt-out and personal-data claims against the data itself. Both models came off public supercomputers. Apertus was trained on CSCS's Alps machine, with more than 10,000 GH200 GPUs ^[17], and EuroLLM on MareNostrum 5 ^[18].

The performance claims come from the builders. The consortium says EuroLLM outperforms similar-sized models on EU multilingual benchmarks and machine translation ^[8]. Apertus's developers report that it significantly outperforms leading closed and open models on rare and regional languages, Romansh and Swiss German among them ^[9]. Those are results on the builders' test sets. They carry over to a production workload when three conditions hold. The traffic has to be in the languages measured. The task has to look like translation or a benchmark prompt. The variant Workers AI serves has to be the size that was tested. The third condition matters most for EuroLLM, because its claim is relative to similar-sized models ^[8].