We Got Time on Leonardo and LUMI. Here's What We're Building.

Two weeks ago we logged into two terminals that looked very different from our usual development machines. One greeted us with an ASCII art "LEONARDO" banner and a SLURM queue. The other, an ASCII art "LUMI" rendered in pixel-block letters, described itself as "The Supercomputer of the North."

Both are EuroHPC Joint Undertaking systems. Both are among the most powerful computers on the continent. And both are now part of how we build EULLM.

What We Got

Leonardo is hosted at Cineca in Bologna, Italy. The booster module we have access to runs 3,456 compute nodes, each with 32 Ice Lake cores at 2.60 GHz, 512 GB of RAM, and four NVIDIA Ampere A100 GPUs with 64 GB of VRAM each. The interconnect is 200G HDR Infiniband Dragonfly+. It is, in short, a serious machine.

LUMI is hosted at CSC in Kajaani, Finland. It calls itself "The Supercomputer of the North," and by the numbers it earns the title. Critically for us, LUMI's GPU partition is built entirely on AMD hardware — which means ROCm, not CUDA.

What We're Doing on Leonardo

The main work on Leonardo is our first vertical model: Legal IT.

The goal is a 7B specialist trained on curated Italian case law, the civil code, and the EU regulatory corpus — a model that understands contracts, GDPR assessments, and EU compliance questions in a way that a general-purpose 70B cannot. Not because it's bigger, but because every parameter is relevant.

The pipeline is: a frontier 70B open-weight model as teacher, structural pruning to remove capacity that's irrelevant to legal reasoning, knowledge distillation into the 7B student, quantization for inference efficiency, and identity fine-tuning for persona and instructions. The Forge pipeline, in other words, run on real iron.

Before we got to training, we wanted to verify that EULLM Engine could actually drive the hardware. So we ran a 27B model across all four A100s on a single node. It worked. The multi-GPU inference path — CUDA-based, using EULLM's own batching and memory management — handled a 27B GGUF without issues, with the expected throughput gains from distributing the model weights across four 64 GB cards.

That's a meaningful data point. It tells us that when the Legal IT 7B ships, a European organization with a single 4×A100 node (on-premise or cloud) will be able to run it comfortably — with room to serve concurrent users.

What We're Doing on LUMI

LUMI is a different challenge. LUMI's GPU partition runs AMD Instinct accelerators, which means the relevant stack is ROCm, not CUDA.

This matters more than it might seem. "Runs on GPU" is not a monolith. Code that works perfectly on NVIDIA cards may not compile, may crash, or may run with poor performance on AMD hardware. EULLM Engine already ships with ROCm support — we marked it as experimental in the release notes — but experimental is not validated.

LUMI is where we validate it properly. The goal is to run EULLM's inference stack across multiple AMD GPUs and confirm that throughput, correctness, and stability hold up under real load. We're running the same model weights, the same API surface, on radically different silicon.

Why does this matter for European AI sovereignty? Because hardware lock-in is a form of dependency. If EULLM only works reliably on NVIDIA, then European organizations running AMD infrastructure (and there are significant AMD deployments across European HPC and research networks) are left out. LUMI gives us a real multi-GPU AMD environment to work against rather than a simulator or a single-card test machine.

Why EuroHPC

We could have rented GPU time from a US hyperscaler. We didn't.

EuroHPC systems are European public infrastructure. The data we compute on stays within European jurisdiction. The allocation process is competitive and merit-based — you submit a project description, you explain what you're building and why it needs the compute, and a committee decides. We think there's something right about building a European sovereign AI stack on European sovereign compute.

There's also a practical argument: the hardware at Leonardo and LUMI is genuinely excellent. Four A100 64GB cards per node is not something you can easily replicate on spot instances. LUMI's scale of AMD GPUs is simply not available for rent anywhere else in Europe.

What Comes Next

Legal IT is a pipeline in progress. The dataset curation is nearly complete. The pruning and distillation runs will happen over the next weeks on Leonardo. When the 7B model passes our internal quality thresholds — and the AI Act compliance card is drafted — it will go into the Hub.

For LUMI: ROCm multi-GPU validation runs are ongoing. We will publish a technical note with benchmarks when the results are solid enough to stand behind.

If you're building on European infrastructure and want to know when Legal IT is available, the best place to follow is our GitHub. Issues, discussions, and release notes all happen there in public.

We're building this in the open, on European machines, for European use cases. That's the whole point.