Europe trains a trillion-parameter model on public iron
Domyn says it trained its EUROPA model at trillion-parameter scale on the JUPITER supercomputer in Germany, with throughput up 14-fold, while a German district pilot of sovereign AI for care cut processing time by a fifth.
- Domyn has demonstrated it can train the EUROPA model at the scale of one trillion parameters, the company says. The EUROPA model was trained at approximately 952 billion parameters. Self-reported m:…0280
- The team increased training throughput 14-fold, from 90 to approximately 1,300 tokens per second per GPU. Self-reported m:…0280
- The model is a Mixture-of-Experts language model with about 1T total parameters, trained on 384 GH200 superchips across 96 nodes, as part of the European Frontier AI Grand Challenge. Self-reported m:…2204
- The Borken district pilot of AI-supported processing of long-term care applications achieved a 35 percent time reduction for selected work steps and a 20 percent time reduction across the entire application process. Self-reported m:…1007
A trillion parameters on a public supercomputer
Domyn has demonstrated it can train the EUROPA model at the scale of one trillion parameters, and the EUROPA model was trained at approximately 952 billion parameters. The team used the JUPITER supercomputer in Jülich, Germany, for training. The point is not just the size. It is where the training happened: on a public European supercomputer, not in a private data centre owned by a US lab. m:…0280
The team increased training throughput 14-fold, from 90 to approximately 1,300 tokens per second per GPU. Throughput increased from 90 to about 1,300 tokens per second per GPU, according to the company's engineering account, after the team rebuilt the training stack around PyTorch and TorchTitan. The team contributed fixes to PyTorch and TorchTitan to benefit other developers, which is the kind of upstream contribution that makes a public training run useful beyond the model itself. m:…0280 m:…2204
Domyn is training a frontier-scale model as part of the European Frontier AI Grand Challenge. The EUROPA consortium includes Fraunhofer IAIS and Fraunhofer IIS, and the consortium has access to 2.5% of the EuroHPC compute capacity. The team is using the Jupiter supercomputer at Forschungszentrum Jülich for the pretraining run. A 2.5% share of EuroHPC capacity sounds small until you remember what it buys: a frontier-scale training run on hardware the consortium does not have to own. m:…2204
The model is a Mixture-of-Experts language model with about 1T total parameters, and the training configuration uses 384 GH200 superchips across 96 nodes. The team adopted the Muon optimizer for weight matrices, and activation checkpointing and offloading to Grace memory enabled longer sequence lengths. The project aims to create an open frontier model with released weights and training recipes. Released weights and recipes are the commitment that makes this a sovereignty play rather than a press release: Europe would own not just the model but the knowledge of how to build it. m:…2204
Sovereign AI reaches the town hall
Materna and deepset are partnering to bring AI applications for administration into secure and sustainable production operation. The Borken district pilot project tested AI-supported processing of applications for long-term care assistance. This is the other end of sovereign AI: not a frontier model, but a local government office processing care applications with a system it controls. m:…1007
The Borken district pilot achieved a 35 percent time reduction for selected work steps and a 20 percent time reduction across the entire application process. The numbers are modest and that is the point. A fifth faster across a whole application process, in a system that has to be correct and auditable, is a real result for a public service, and it came from a pilot, not a demo. m:…1007
KIAN is a digital advisor for care services that is part of the Materna Sinfonie platform, and it uses Retrieval-Augmented Generation to retrieve and incorporate specialist information into answers. Transparent sources in Haystack make AI results verifiable for administrative procedures. The Haystack framework connects language models, knowledge sources and other components for AI applications, and it supports various operating models, including sovereign cloud environments and on-premises data centers. Verifiability is the requirement that separates administration from chat: a citizen denied care needs to see why, and a source link is the answer. m:…1007
Two halves of one bet
Domyn rebuilt the training stack around PyTorch and TorchTitan, and Haystack is the established AI orchestration standard in the German AI stack, according to deepset. The two stories share an architecture of independence: train on public infrastructure with open tools, and run administration AI on infrastructure the state controls. Neither depends on a foreign cloud, and neither asks citizens to trust a black box. m:…0280 m:…1007
The open question is whether the open model arrives and whether the pilot scales. The project aims to create an open frontier model with released weights and training recipes, but aims are not releases, and a district pilot is not a national rollout. Europe has now shown it can train at trillion-parameter scale on public hardware. The next test is whether anyone builds on it. m:…2204
What is still unresolved
- When EUROPA's weights and training recipes will actually be released
- Whether the Borken pilot's time savings hold at national scale
- How much EuroHPC capacity frontier training will need as models grow
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.