Europe's Trillion-Parameter EUROPA: Enough Compute, but Only to Match 2025

What this is about

“EU builds trillion-parameter model”: that's the short version doing the rounds since November 12. On that day, the EUROPA consortium described in the PyTorch blog and in a more detailed companion post on Domyn's site how it got a language model with nearly a trillion parameters running on the JUPITER supercomputer at Forschungszentrum Jülich. The consortium is made up of Milan-based Domyn (formerly iGenius) and the Fraunhofer institutes IAIS and IIS. The European Commission is behind the project: in June, EUROPA won the “Frontier AI Grand Challenge,” and the prize isn't money but compute time.

We ran the consortium's numbers. Our argument: technically, this is a real success, because public research computers can handle training at this scale. But the resulting compute budget lands almost exactly at the level of DeepSeek-V3 and Kimi K2, which date from 2024 and 2025.

What actually happened in Jülich

First, an important caveat: there's no finished model, and the actual training run hadn't started at the time of the report. Domyn writes that the post describes the state of things “at the time of writing” and that some figures may “still change before the start of the training run.” The Italian site Bitmat still ran a headline saying the model had been “trained.” The article itself, though, only talks about a feasibility demonstration “in view of the full execution.”

What was tested is a Mixture-of-Experts model (MoE). The model consists of many specialized subnetworks, and only a few of them are active for each token. The key specs:

The resemblance to Moonshot AI's Kimi K2 is striking. Our take: EUROPA is adopting a proven blueprint rather than trying anything new on the architecture side. For a project with just one year of compute time, that's sensible.

The real work went into adapting to the machine. JUPITER is a research computer shared by hundreds of projects. Each node has four GH200 superchips (each a Hopper GPU with 96 GB plus a Grace CPU with 120 GB of memory). But according to Domyn, the connection between nodes is 36 times slower than the link between a GPU and its own CPU. So the team offloaded intermediate results to CPU memory (activation offloading), moved computations for the Muon optimizer from the CPU to the GPU (an optimizer step then took 2.2 seconds instead of 41.7) and fixed a memory bug that caused allocations to fail despite 25 to 30 GiB of free VRAM. That raised throughput from 90 to about 1,300 tokens per second per GPU, a 14x improvement. Then again, that factor also says a lot about how naive the starting point was.

The team has submitted its fixes to TorchTitan and PyTorch, including hooks for offloading, Muon computations on the GPU, a shared RoPE cache and two PyTorch bug fixes. The work isn't fully done: according to the blog, the key memory option still has to become a public FSDP interface (PyTorch #202343). FSDP spreads the model's pieces across all GPUs.

Running the numbers: what 1,300 tokens per second means

The plan calls for four data-parallel replicas, meaning four copies of the training run that process different data in parallel. Each runs on 384 GPUs, for a total of 1,536 GH200s. Here's what that works out to:

The four months don't come from Domyn itself. Bitmat cites “about four months” as a projection for pretraining, and according to maccus.fr, Quotidiano Nazionale reported the same. Domyn hasn't given a token budget or a date so far.

Now for utilization, often called MFU (Model FLOPs Utilization, the share of theoretical compute that actually goes into useful model computation). The rule of thumb is 6 × active parameters per token, which comes to about 197 billion operations per token. Multiply by 1,293 tokens/s and you get about 254 trillion operations per second. Against a Hopper GPU's peak BF16 performance of 989 TFLOPS, that's about 26%, or roughly a quarter. Domyn itself doesn't state an MFU figure. The calculation underestimates real utilization, because attention over 65,536-token sequences and recomputing intermediate results add to the total. How much they add is impossible to say without details on the attention architecture. For comparison: DeepSeek-V3 was pretrained on 14.8 trillion tokens in 2.664 million H800 GPU hours. That's about 1,540 tokens/s per GPU with 37 billion active parameters, or roughly 35% by the same math. DeepSeek, however, trained in FP8 with a sequence length of just 4,000 tokens.

What “2.5% of EuroHPC capacity” means, and why the size keeps changing

According to the call for proposals, the winner gets “up to 2.5% of total EuroHPC computing capacity for one year,” on one or more AI-capable machines. The EU never published a figure in GPU hours, and applicants had to justify their own needs. Here's our own back-of-the-envelope math: the JUPITER Booster has 5,884 nodes with four GH200s each, or 23,536 GPUs. That's about 206 million GPU hours a year, and 2.5% of that would be about 5.2 million. At 4.4 million, the four-month main run alone would eat up most of that, while occupying 6.5% of the Booster. Count other EuroHPC machines like Leonardo, LUMI and MareNostrum 5 and the budget gets bigger. But ablations, a long-context phase and post-training all have to fit in too. Our take: there isn't enough for a second, larger run.

The shifting size figures have simple explanations. 400 billion parameters was the minimum set by the call, and in June Domyn announced a “400-billion-parameter model within a year.” The test configuration has 951.9 billion, which Domyn rounds to “~1T.” What size the main run will end up being hasn't been officially decided.

Data, timeline, competition

The report says nothing about the data mix: no languages, no proportions. That's exactly where a core problem lies. Data quality provider Pebblous puts Maltese at 0.03% of Common Crawl, Irish at 0.07% and Latvian at 0.09%, though without citing a source. More parameters can't generate sentences that don't exist on the web.

The timeline is tight too. The award came on June 19, 2026, and the compute time is valid for one year. According to Declic Media, Domyn talked about a release “within the year.” Declic quipped at the time that the EU was crowning “the champion of a race that hasn't started yet.” The site listed EUROPA as at least the third parallel sovereignty project.

With Edge, Domyn has at least shown that it can train a model from scratch on public infrastructure (512 A100s on Leonardo). The first planned OpenEuroLLM models aren't due until late 2026, with the final ones coming in January 2028. On the side, Domyn is building a private cluster called “Colosseum” with G42, with around 6,000 Grace Blackwell chips. According to The Next Web, the company has raised more than a billion dollars for it, about 90% of it as debt. Neither blog post describes what exactly the Fraunhofer institutes are contributing.

Our take

The good news first: the team has cleanly documented how to train a trillion-parameter MoE on a machine that was never built for it. It was upfront about weaknesses (group-wise expert routing “can actually hurt training,” and the FP8 check ran for only ten steps) and gave its fixes back to the community. That a run like this is possible on public hardware in Germany wasn't a given before.

Our take on performance: at around 3.6 × 10²⁴ FLOP, EUROPA lands just above DeepSeek-V3 (approx. 3.3 × 10²⁴) and Kimi K2 (approx. 3.0 × 10²⁴) on paper. DeepSeek-V4-Pro used about two and a half times as much. Compute isn't everything, since data and post-training matter too. But barring a miracle on the data side, the realistic target is the level of Chinese open-weight models from 2025, and with a release in mid-2027 that would be about two years old. That can still be valuable for European agencies and companies looking for an open, documented model built for all 24 official EU languages. But EUROPA won't compete with the world's best. Still unresolved: the license, the token budget, the data mix, the start date of the main run, and who will maintain the model once the compute time runs out.