EuroHPC: Cutting-Edge Research and AI with JUPITER
The JUPITER supercomputer is procured through the EuroHPC Joint Undertaking, a public-private partnership to advance High-Performance Computing (HPC), in the European Union. EuroHPC manages 50% of the compute time of the system, the other 50% are awarded by the Gauss Centre for Supercomputing (GCS) to German universities and research institutions.
EuroHPC-JUREAP
Similar to GCS, EuroHPC has also selected Lighthouse Projects as part of the JUPITER Research and Early Access Program (JUREAP) and allocated significant computing time on JUPITER to them. This computing time will be used after JUPITER has gone into full operation to provide timely answers to critical questions of our time. Until then, the 15 selected JUREAP-EuroHPC Lighthouse Projects will be prepared for JUPITER in collaboration with JSC. The selected projects have a focus on AI, but also include other important scientific areas, such as climate research, fluid mechanics and particle physics.
The JUPITER AI Factory (JAIF)
In March 2025, EuroHPC also awarded the JUPITER AI Factory, led by JSC, which uses JUPITER as a central element. The AI Factories offer a wide range of services for startups, SMEs, industry and the scientific community, and are an ideal framework for some EuroHPC-JUREAP Lighthouse Projects. Ten of the EuroHPC-JUREAP Lighthouse projects are led by academic partners, five by startups/SMEs. This underlines the ambition to make European supercomputers more accessible to startups, SMEs and industries in the future, in order to further promote European innovation.
JUPITER USE CASES
Jean-Gabriel Barthelemy, Machine Learning Engineer at Lingua Custodia, France
Our goal is to build frugal and highly efficient multilingual LLMs for European languages. Our research focuses on improving training efficiency (learn faster) and text generation throughput, being able to run LLMs with less energy and less powerful machines. With JUPITER, we gain unprecedented computational power to train models faster and at a greater scale. By ensuring open access and compliance with the EU AI Act, we aim to democratize AI, fostering innovation and bridging linguistic gaps for a more inclusive digital future.
Dr. Toms Bergmanis, lead AI researcher at Tilde, Latvia
We are developing TildeLM, an open foundational large language model with over 30 billion parameters, designed to cover all European languages, with a special focus on Baltic and Eastern European languages. The computational power of JUPITER will enable us to leverage open-weight models to synthesize new training data and adapt TildeLM for downstream applications in language services such as translation, summarization and retrieval-augmented generation. Our goal is to provide high-quality, culturally relevant AI-driven language technologies for 250 million users across Europe and beyond. These efforts will facilitate Europe's digital sovereignty and drive innovation in cross-border communication.
Prof. Dr. Matteo Bernardini, Sapienza Università di Roma, Italy
Our research explores the physics of high-speed turbulent flows, with the main goal of understanding aerodynamic heating and drag in supersonic and hypersonic vehicles. Through large-scale direct numerical simulations, we study complex shock wave-boundary layer interactions and heat transfer mechanisms on both smooth and rough surfaces. With JUPITER, we plan to leverage Exascale computing to capture turbulence at an unprecedented level of detail, opening new avenues for scientific discovery. This innovative approach not only enhances our understanding of fluid dynamics but also drives advancements in aerospace safety and efficiency, ultimately benefiting society through safer, more sustainable high-speed travel.
Prof. Dr. Alexandra Birch, University of Edinburgh, UK
We are researching multilingual large language models that cover the needs of European citizens. We are focussing on creating models that can reason in any language and can reason over long documents. JUPITER will help us to train these models with large amounts of multilingual data and to create large synthetic training sets of long documents and reasoning chains to unlock previously unachievable performance. This will lead to models which are useful for both researchers and industry and government having access to the best Generative AI models that address their needs across borders in the EU.
Dr. Robin Kurtz, Senior Data Scientist at KBLab, National Library of Sweden
In the project 'Scandinavian Language Models', we train Large Language Models (LLM) for Swedish, Norwegian, and Danish. Given the vast amount of compute and data necessary, we focus on continued pretraining of existing models, adding or increasing the models' language capabilities for the three Scandinavian languages. Access to JUPITER's increased compute allows us to scale up the training process to larger models and larger datasets. The most performative LLMs are optimized for English, and are only accessible through expensive APIs. Publishing open models thus helps in the democratization of AI, especially for lower resourced languages such as Swedish, Norwegian, and Danish.
Prof. Dr. André Martins, University of Lisbon, Chief Science Officer at Unbabel, Portugal
Our research aims to develop the next generation of AI foundation models for language generation, equipped with advanced cognitive capabilities. By integrating concepts from machine learning, sparse modeling, information theory, and cognitive science, we adopt an interdisciplinary approach to push the boundaries of multimodal reasoning. Through JUPITER, we strive to scale multimodal and multilingual language models that support all European languages, addressing the limitations of existing models. The models trained on JUPITER will be openly released, benefiting the entire research community and European companies seeking to build advanced AI systems.
Dr. David Montero, Multiverse Computing, Spain
We are researching novel model compression techniques to democratize access to powerful large language models. Our current focus is Deepseek-R1 (671B parameters), aiming to drastically reduce its size while preserving accuracy and eliminating embedded political bias. With JUPITER's unprecedented AI compute capabilities, we will apply advanced tensor network methods to enable extreme compression with minimal quality loss. Exascale resources are essential to process models of this scale and complexity efficiently. This work empowers broader, unbiased use of AI while drastically reducing environmental and infrastructure costs—delivering real societal impact in accessibility, fairness, and sustainability of future foundation models.
Dr. Simon Šuster, NLP Scientist at Textgain, Belgium
We build on the CaLICO project, which sets the stage by creating the world’s first foundation model capable of processing harmful content. Our current focus is on developing advanced text encoders, which are non-generative in nature and efficiently create contextualised representations for AI applications such as document retrieval and text classification. With JUPITER, an exascale supercomputer, we can scale training to massive multilingual datasets, including previously untapped sources like social media. By parallelising computation across hundreds of GPU nodes, we accelerate development cycles and contribute to advancing language technology.
Contact
- Institute for Advanced Simulation (IAS)
- Jülich Supercomputing Centre (JSC)
Room 307
Dr. Andreas Herten
Co-Lead of division Novel System Architecture design, head of ATML Accelerating Devices PI in Helmholtz Information Program 1, Topics 1 and 2
- Institute for Advanced Simulation (IAS)
- Jülich Supercomputing Centre (JSC)
Room 228