EUROPA shares a trillion-scale training recipe, not a finished frontier model
EUROPA shares a trillion-scale training recipe, not a finished frontier model
Event/source date: 2026-11-12 technical reports; experiment execution dates unknown
The EUROPA consortium has published an engineering account of training a near-trillion-parameter mixture-of-experts configuration on JUPITER, Europe’s shared public supercomputer. The November12 report describes a 951.9-billion-parameter experimental configuration with 32.8 billion active per token. It does not announce downloadable finished model weights. PyTorch report
The team reports raising throughput from 90 to roughly 1,300 tokens per second per GPU on 384 GH200 GPUs. Its changes address optimizer communication, activation offloading and memory pressure. These are the consortium’s measurements, not independent benchmarks.
The longer Domyn account supplies an important qualification: the sequence length changes between the initial and final configurations. Its chart uses steady-state measurements from steps four through ten, and describes a changing setup. The headline improvement is therefore not a fixed-configuration comparison of one isolated optimization. Technical account
Analysis: make the machine fit the workload
The useful story is less “Europe has released a trillion-parameter model” than “public research hardware can require a different training recipe.” Raw accelerator specifications do not explain how much time a job spends waiting for data or how a shared machine limits an engineer’s options.
For public-compute operators, this is a concrete case study to investigate rather than a promise to copy the reported speedup. For prospective model users, it changes no deployment decision yet: they still need the released artifact, license, evaluation and serving requirements.
What remains unproven
This run did not test the training setup or independently establish the reported throughput. The report is not evidence of completed pretraining, released weights, frontier task quality or legal compliance. The source itself says experimental values can change. The actual experiment dates are not specified; November12 is the publication date.
Cited primary evidence
- EUROPA consortium report on PyTorch: 2026-11-12. Directly read; configuration and attributed throughput claim
- Domyn detailed engineering account: 2026-11-12. Directly read; changing sequence lengths, measurement scope and evolving setup; same research team, not independent
Independent evidence
No independent reproduction verified in this run. Search discovery summaries were not treated as replication.
Sources and citations
Revision history
Published .