Europe’s AI Strategy: Small Models, Big Impact
Europe is charting a unique course in the global artificial intelligence (AI) race, moving beyond the singular pursuit of massive language models and embracing a strategy centered on specialized, smaller AI models (SLMs). This approach, fueled by initiatives like the OpenEuroLLM project launched in February, aims to foster digital sovereignty, boost competitiveness, and democratize access to AI growth within the continent.
The European landscape differs significantly from the U.S., where giants like Google and OpenAI dominate the AI conversation with their massive language models (LLMs). While European companies like Mistral AI (France) and Aleph Alpha (Germany) have been considered potential competitors,the emphasis is shifting towards a more sustainable and targeted approach. Aleph Alpha,for example,is now concentrating on providing AI infrastructure and platforms for enterprise and government clients,rather than solely focusing on foundational llms.
the OpenEuroLLM Project: A Collaborative foundation
OpenEuroLLM embodies this collaborative spirit. This consortium, comprised of 20 leading European research institutions, companies, and EuroHPC centers, is constructing a family of high-performing, multilingual, large language foundation models. the project,coordinated by Jan Hajič of Charles University (Czechia) and co-led by Peter Sarlin of AMD Silo AI (Finland),distinguishes itself by its commitment to Europe’s rigorous regulatory framework,aligning with European values,and fostering open-source collaboration. The goal is to ensure these models, along with their associated software, data, and evaluations, are fully open, enabling fine-tuning and instruction-tuning for specific industry and public sector needs – a move that could resonate well within the U.S. government agencies interested in open-source solutions.
Though, a key question remains: can Europe close the AI innovation gap by focusing on smaller, specialized AI models and applications?
The Case for Small Language Models (SLMs)
anita Schjøll Abildgaard, CEO of EU-funded Iris.ai, champions a future for European AI that relies on smaller, domain-specific models (SLMs) and open-source collaboration. This approach addresses a critical challenge: the escalating energy demands of data centers. Estimates project a tripling of data center power usage in Europe by 2030, making energy-efficient AI solutions a necessity.
“With Europe’s data center power demand set to triple by 2030, it’s compelling to envision a different approach to AI — one that doesn’t rely on ever-larger models consuming vast energy resources,” Abildgaard argues.SLMs offer a compelling alternative. While they require foundational models to generate the high-quality data needed for training, they are significantly cheaper, can be tailored to specific business needs, and are frequently enough more practical than general-purpose LLMs.
Victor Botev, CTO and co-founder of Iris.ai, emphasized that most business use cases simply don’t require the immense power of massive LLMs. “If you can distill just the knowledge you need into a smaller model,it’s more efficient and cost-effective,” Botev stated. “That’s been our focus — using small models to address real workflows.”
SLMs find applications in scenarios demanding fast, efficient decision-making, such as agent-based workflows and specialized domains like chemistry and healthcare. According to Botev,”Training large models to understand DNA or chemical structures risks ‘catastrophic forgetting’ — where they lose existing capabilities. Small models let us specialise without that risk.”
The energy savings are substantial. “A 1B parameter model needs far less compute than a 400B one — perhaps 60,000x fewer resources,” Botev pointed out. “Small models can run on older GPUs with lower power use. They’re not only cheaper but greener too, wich aligns with Europe’s sustainability goals.”
Openness and Collaboration: Europe’s Competitive Edge
europe sees open-source collaboration as a key driver of innovation. As Botev notes, “They’ve open-sourced distillation methods, which let people build small models from large ones. There are now over 10,000 distilled models on Hugging Face.” He also highlights the effectiveness of reinforcement learning in customization, even rivaling supervised fine-tuning. “However in Europe have expertise in safety, guardrails, and efficient systems — we should double down on that.”
Abildgaard suggests that European startups should consider adopting principles from open-access publishing. “If you get EU funding, maybe you should be required to open-source some of your work, especially foundational models,” she proposes. “It would drive collaboration and transparency — areas where Europe leads.” This approach contrasts with the more proprietary models frequently enough associated with U.S. AI development and could appeal to the U.S. government’s increasing interest in transparency and ethical AI deployment.
Iris.ai’s partnership with Sigma2 AS in Norway exemplifies this collaborative approach. Sigma2 provides the national e-infrastructure for computational science, offering high-performance computing and large-scale data storage for research and education. Iris.ai uses Sigma2 to train and domain-adapt small models (1–9B parameters) and evaluate system components. According to Botov, evaluation is often overlooked, “but it’s compute-intensive and critical for scaling systems with multiple agents and retrieval layers.”
RAG Systems and the Future of SLMs
Iris recently launched a new business line – a powerful Retrieval-Augmented Generation (RAG) system,Iris-ai’s RAG system is agent-based and packed with small models. This system exemplifies the potential of SLMs in real-world applications.
Despite these advancements, SLMs and local competitors to LLMs have been slow to emerge in Europe. One notable exception is Malted AI (Scotland), which distills the output of large models into smaller, domain-specific models, allowing enterprises to achieve significant cost savings.
U.S. Implications and the Path Forward
Europe’s strategy of focusing on SLMs and open-source collaboration presents a viable alternative to the U.S.’s current LLM-dominated landscape. The emphasis on energy efficiency and specialized applications could be particularly relevant as the U.S. grapples with its own growing energy demands and seeks AI solutions for specific industries and government functions.
Rather than solely pursuing scale,Europe is aiming to lead by championing smart,open,and sustainable AI. this entails investing not only in foundational models but also in the ecosystems that foster agile development, collaboration, and fine-tuning. Ultimately, the future of European AI might lie not in building the biggest model but in creating the smartest, most specialized ones. It also resonates with concerns about AI’s carbon footprint, and the need for sustainable development. This approach could offer valuable lessons for the U.S.as it navigates the evolving AI landscape.
The success of this European model will depend on continued investment, collaboration, and a focus on translating research into practical applications. The results of the OpenEuroLLM project, coupled with the innovative work of companies like Iris.ai and Malted AI, will be crucial in determining whether Europe can carve out a unique and impactful position in the global AI arena.
Considering Europe’s focus on smaller, specialized AI models, how might this approach differ in its impact on job markets compared to the widespread adoption of large language models?
Interview: Europe’s AI Strategy – Focusing on Small Models for Big Impact
Archyde News: Welcome, Dr. Elena Petrova, Lead AI Researcher at the european Institute for Technological Innovation. Thank you for joining us today to discuss Europe’s innovative approach to AI development. The continent seems to be prioritizing Small Language Models (SLMs) over the massive ones dominating headlines. Can you shed some light on this strategy?
dr. Petrova: Thank you for having me. Yes, Europe is indeed taking a distinct path. While the US and other regions are heavily invested in developing extremely large language models (llms), we are focusing on SLMs. This move stems from a variety of factors, including our commitment to digital sovereignty, sustainability, and promoting broader access to AI technologies. We believe that SLMs offer a more efficient and practical approach for manny applications.
The openeurollm Project and Collaborative Spirit
archyde News: The OpenEuroLLM project seems to be a cornerstone of this strategy. Can you explain its goals and how it fosters collaboration?
Dr. Petrova: OpenEuroLLM embodies the collaborative spirit at the heart of our strategy. It is indeed a consortium of over 20 leading European research institutions and companies, working together to create a suite of high-performing, multilingual large language foundation models. The models and associated resources such as data, and evaluations will be open source, allowing for fine-tuning and customization for diverse industry and public sector needs.
archyde News: The sustainability aspect is also often highlighted. Data centers’ power demand is projected to increase. Can you detail more on how SLMs contribute to more lasting AI?
Dr. Petrova: Absolutely. SLMs offer a significant advantage when it comes to energy efficiency. Training and running these models require far less computing power than massive LLMs. This is crucial, given the projected increase in data center energy consumption in Europe. More efficient models translate directly into reduced carbon footprint, aligning with our sustainability goals.
Real-World Applications and Advantages of SLMs
Archyde News: What are some specific applications where SLMs excel, and how do they differ from LLMs in these scenarios?
Dr. Petrova: SLMs are notably well-suited for scenarios demanding fast, efficient decision-making such as agent-based workflows, or specialized domains like chemistry or healthcare. Unlike LLMs, SLMs can be specifically trained and adapted for well-defined tasks, enabling greater efficiency and cost-effectiveness. Furthermore, SLMs can be more easily customized.
Archyde News: Open-source collaboration seems to be a key element of Europe’s approach. How does this open and collaborative surroundings contribute to innovation?
Dr. Petrova: Open-source promotes innovation by allowing the community to build on each other’s work, accelerating development. It encourages openness and sharing of best practices, and avoids the development of highly proprietary models and the potential for lock-in by dominant players. In Europe we have a strong interest in safety, ethical development, and this mindset is vital for building trust.
Challenges and the Path Forward
Archyde News: Despite these advantages, are there any unique challenges to European AI development?
Dr. Petrova: yes. SLMs and local competitors to LLMs have been slow to emerge in Europe. A key challenge is attracting investment and talent.
Archyde News: Europe’s strategy offers a potentially valuable alternative to the current LLM-dominated approach. How do you see this evolving and impacting the global AI landscape, especially with U.S. interests?
Dr. Petrova: I believe it could provide valuable lessons.The emphasis on energy efficiency, specialized models, ethical use, and open collaboration could influence future AI strategies. We are moving to a model that focuses on the smartest AI and not necessarily the biggest,and how the results of our projects affect the global AI landscape will be very engaging to see.
Archyde News: Dr.Petrova, this has been an insightful discussion. Thank you for your time and expertise. Could you provide one final thought to our readers?
Dr. Petrova: I would encourage our readers to consider the long-term implications of AI development. Is bigger always better? Can specialized,smaller models offer a more viable,sustainable,and ethical path forward? I invite readers to share their thoughts on this topic.
