Clarity for AI buildersSponsored by AI Routing
availability

Domyn tests 951.9-billion-parameter EUROPA model on Europe’s JUPITER supercomputer

Domyn tests 951.9-billion-parameter EUROPA model on Europe’s JUPITER supercomputer

Domyn has tested a 951.9-billion-parameter EUROPA model on JUPITER, but has not released its weights, MACCUS reports.

The trials demonstrate an approach to training a large AI model on shared European public computing infrastructure, rather than a completed model release. According to MACCUS’s account of Domyn’s November 12, 2026 technical post, engineering changes increased training throughput from 90 to 1,293 tokens per second per GPU—roughly a fourteenfold improvement. The measurements do not establish the model’s eventual capabilities or confirm completion of its main training run.

EUROPA is led by Milan-based Domyn, formerly iGenius, with Germany’s Fraunhofer IAIS and IIS as partners. MACCUS reports that the European Commission selected the consortium on June 19 through the Frontier AI Grand Challenge. The competition required a model with at least 400 billion parameters and offered up to 2.5% of EuroHPC’s total computing capacity for one year.

The project must cover all 24 official EU languages.

A large model with a smaller active footprint

The trial configuration has 951.9 billion parameters, more than twice the Commission’s minimum. It uses a mixture-of-experts architecture: eight of its 384 specialized subnetworks are selected for each token, leaving 32.8 billion parameters active per token. Training sequences contain 65,536 tokens. Domyn cautions that some configuration values may change before the full training run.

That distinction matters when interpreting the headline size. The total parameter count describes the model’s overall scale, while the active count helps explain how its computation is limited for each token. Neither figure, by itself, demonstrates answer quality, multilingual coverage or performance against other models. The supplied report contains no benchmark results for EUROPA.

Working around a shared supercomputer’s limits

JUPITER, located in Jülich, Germany, serves scientific projects as well as AI workloads. Its Booster module contains 5,884 nodes, each with four NVIDIA GH200 chips. According to MACCUS’s account, Domyn has user-level access and cannot reconfigure the operating system, network or node memory settings.

The main obstacle was communication between nodes. Each model copy spans 384 GPUs across 96 nodes, which must exchange training data. Domyn reported that the connection between nodes was 36 times slower than the connection between a GPU and the CPU on its own chip.

The team lengthened training sequences to give GPUs more work between exchanges and moved some intermediate results into CPU memory to free GPU memory. It also shifted the most demanding part of the Muon optimizer from CPU to GPU, reducing its time per training step from 41.7 seconds to 2.2 seconds. Other changes included transmitting expert weights in eight-bit FP8 format and gradients in sixteen-bit bf16 format.

A memory-management fix raised throughput by 21%.

Taken together, these changes address the practical problem of keeping processors busy when communication is constrained. The result is evidence about training infrastructure and software efficiency—not yet evidence that EUROPA will meet its intended users’ needs.

Short measurements leave important questions open

MACCUS reports that throughput figures were medians measured over steps four through ten of each trial. Domyn estimated that four parallel model copies running on 1,536 GPUs could process about 150 billion tokens a day, retaining 89% of the per-GPU speed measured with one copy. Those short trials do not establish sustained performance across a complete training run.

There are also unresolved quality trade-offs. Domyn says a routing rule that distributes tokens evenly across groups of experts can hurt training. Its finding that compressed weights and gradients did not measurably affect loss was based on comparisons covering only ten training steps.

The remaining milestones therefore go beyond scaling the hardware. Although Domyn plans to publish EUROPA’s weights and training-method details, the supplied report gives no release date, license, training-data description, language proportions or evaluation scores. Until those are available, the trials support a narrower conclusion: Domyn has demonstrated substantially faster execution of a near-trillion-parameter training configuration on JUPITER, not delivered a finished open model.