Paris, October 15th, 2025 – Today, Dragon LLM — formerly known as Lingua Custodia — announces the launch of Dragon, a groundbreaking AI architecture developed and trained on EuroHPC’s European supercomputers (Leonardo, JUPITER). Dragon LLM was recognised as a winner of the European Large AI Grand Challenge in Ju
ne 2024, securing access to the EU’s extensive computing resources, which made this major technological achievement possible.
Dragon represents a new generation of European artificial intelligence: efficient, accessible, resource-conscious, and fully sovereign.
The Dragon AI architecture, designed to power future large language models, redefines performance and sustainability benchmarks for language models. Unlike current leading models (such as GPT, Claude, and Llama), which rely on the energy-intensive Transformer architecture, Dragon employs a hybrid design optimised for long-form text and complex reasoning. This innovation significantly reduces energy consumption and inference costs—while maintaining equivalent or superior output quality.
A European first for business
Dragon is built on the successful integration of cutting-edge research from leading laboratories worldwide over the past two years. This breakthrough architecture represents a true game changer — a streamlined AI that trains faster, costs significantly less to operate, and is designed for simple, sustainable use by organizations of all sizes, from large enterprises to SMEs.
This landmark achievement marks a turning point for Europe’s AI ecosystem. It demonstrates that innovative startups and SMEs across the continent have access to the computing power and talent needed to compete with the United States and China in developing revolutionary AI architectures and models.
Olivier Debeugny, CEO and founder of Dragon LLM, explains his vision: “Our goal is simple — to build responsible and accessible AI for European companies that delivers real return on investment. Europe can innovate and create sovereign models, and Dragon is proof of that. It’s a powerful symbol: a small French Tech company developing AI models that rival the performance of major international players — without raising any external funding. Dragon shows that Europe can build and own its own AI technologies.”
What Dragon Demonstrates
- Reduced computational requirements: Dragon achieves the same level of performance as other leading models while using only one-third of the computational power.
- Standard server deployment: The model can be deployed on conventional servers without the need for vast GPU resources.
- Exceptional energy efficiency: At a time when European data centers are facing a threefold increase in energy consumption, Dragon offers a sustainable alternative. Training is faster, requires less data, and minimizes resource use before the model even starts generating value.
- Outstanding operational performance: With equivalent hardware, Dragon can serve twice as many users as competing models, while matching the results of leading Chinese (e.g., Qwen) and US (e.g., Meta) models.
The first Dragon model architecture is built on 3.6 billion parameters, with larger versions to be announced in the coming months.
The Dragon Architecture Code and Demonstration Model are publicly available on the Hugging Face platform ; https://huggingface.co/DragonLLM/Dragon-3B-Base-alpha
| About Dragon LLM Founded in 2011 under the name Lingua Custodia to target the financial industry, Dragon LLM is a French company specialised in frugal and sovereign AI models & architectures. Winner of the Large AI Grand Challenge organised by the European Commission, Dragon LLM designs efficient and economical AI models, adapted to business practices and deployed on local infrastructures. Based in Paris, the company aims to build a useful, responsible and efficient European AI, at the service of companies while promoting the technological sovereignty of the continent. https://dragonllm.ai/ |