EuLLM: The European Platform for Sovereign LLMs That Challenges ChatGPT

Every time an organization sends a document to ChatGPT or any other American LLM, that text crosses the Atlantic, gets processed on servers outside the EU, and — in most cases — contributes to training future models. For a confidential contract, a medical record, or a financial risk assessment, this isn't a technical detail: it's a serious legal problem.

EuLLM was built to solve exactly this problem. It's not yet another wrapper around GPT-4 hosted on an Amazon European datacenter. It's a platform built entirely in Europe — servers, models, team, and code — that brings generative AI inside your own infrastructure without data ever leaving the European Union.

Why "European" isn't enough if the model is American

There's an important distinction that often gets ignored in AI marketing: using an AWS European server doesn't solve the sovereignty problem. The value chain remains American — from the chip to the model, from the code to the company that develops it. The Schrems II ruling of 2020 and the EU AI Act that entered into force in 2024 have made this problem very real for anyone working in regulated sectors.

EuLLM responds with a fully European and auditable stack: Qdrant for the vector database (Berlin), Mistral AI for base models (Paris), I3K Technologies for platform development (Milan). Zero non-EU dependencies in the path of a single request.

How it works: turning a large model into a small expert

EuLLM's core idea is vertical distillation. You take a 70-billion-parameter generalist model and transform it into a 7-billion-parameter specialist focused on a specific domain — law, medicine, finance. The result is not a worse model: within its field, the 7B outperforms the original 70B.

How is this possible? A generalist model distributes its capacity across dozens of domains. Most parameters are "wasted" for any specific use case. Distillation eliminates this waste. You get a more accurate model, much faster, that runs on hardware you already own — no need for enterprise GPUs costing tens of thousands of euros.

The platform consists of three components working together:

Forge is the foundry where transformation happens. The pipeline includes structural pruning (removing parameters irrelevant to the domain without retraining from scratch), knowledge distillation, quantization to maximize throughput, and identity fine-tuning to customize persona and brand voice. The output is a GGUF file ready for production.

Hub is the catalog of pre-specialized models, hosted entirely in the EU. Each model is distributed with an AI Act compliance card — a document certifying risk category, training dataset, and usage limitations. The first vertical arrives in Q3 2026: Legal IT, trained on Italian case law, the civil code, and European regulatory corpus. Medical and Finance will follow by end of 2026.

Engine is the inference server written in Rust. It replaces Ollama with an OpenAI-compatible API, reaches 259 tokens per second across 16 concurrent requests, handles multimodal input (images and audio) entirely locally, and generates no telemetry to non-EU servers. Supports NVIDIA CUDA, AMD ROCm, Vulkan, and Apple Metal.

2026 is the year the AI Act becomes operational

Those working in legal, healthcare, or financial fields already know that AI systems used in these sectors are classified as "high risk" by the EU AI Act. This means concrete obligations: technical documentation, conformity assessment, usage registry, human oversight.

The compliance cards integrated into every Hub model are the technical response to these obligations. They're not an accessory — they're what a DPO or compliance officer can show during an inspection.

EuLLM is a candidate for the EIC Accelerator 2026 under Horizon Europe, the EU program for deep tech with continental market potential.

Who should evaluate it

If your organization cannot send confidential documents to American APIs — for legal, contractual, or internal policy reasons — EuLLM is likely the most complete solution available in Europe today. The code is open-source (Apache 2.0) and available on GitHub, so it can be verified by anyone with the technical skills to do so.

For IT infrastructure managers, the most relevant data point is that Engine runs on hardware you already have: you don't need to buy anything new to get started.

Official website: eullm.eu — Source code: github.com/eullm/eullm