EU open model EUROPA tests 1T scale
Domyn and the European Commission ran pretraining trials for EUROPA at a 952-billion parameter Mixture-of-Experts scale on Germany's JUPITER exascale supercomputer. The open-weights model activates 32.8 billion parameters per token and targets native coverage across all 24 official EU languages.
Europe is demonstrating that public supercomputing infrastructure can realistically train frontier-scale open models without reliance on commercial US clouds.
- –The 952B architecture utilizes 384 routed experts with top-8 routing, keeping per-token inference demands at a manageable ~32.8B active parameters.
- –Adapting TorchTitan and PyTorch to NVIDIA GH200 nodes successfully converted a simulation-oriented exascale cluster into an efficient AI training platform.
- –Shifting matrix parameter optimizations to the Muon optimizer while offloading activation checkpoints to the Grace CPU helped bypass inter-node network bottlenecks.
- –If delivered as true open weights, EUROPA provides a sovereign foundation for European institutions and developers competing with proprietary closed ecosystems.
DISCOVERED
9h ago
2026-11-16
PUBLISHED
9h ago
2026-11-16
RELEVANCE
AUTHOR
IntCyberDigest