From policy to practice: how ALIA strengthens Spain’s public AI infrastructure

As governments across the world look to harness the potential of artificial intelligence, many are investing in large-scale infrastructure to develop sovereign AI systems and ensure national languages are represented in these technologies. By building these capabilities, countries aim to strengthen their ability to adopt AI effectively while safeguarding their cultural and linguistic diversity.

In Spain, ALIA is emerging as a significant step in this direction. For our most recent Spotlight Series entry, we spoke with Josep M. Martorell, Associate Director of the Barcelona Supercomputing Center (BSC), to learn more about ALIA’s objectives, the challenges it aims to address, and its potential impact.

ALIA has gained attention within the AI policy community as Europe’s first public, open, and multilingual Large Language Model (LLM). However, ALIA’s scope extends beyond being just an LLM. According to Martorell, ALIA is a public infrastructure designed to support AI development in Spain while promoting the use of Spanish and the country’s co-official languages. ALIA’s infrastructure is built on three core elements:

• The existence of language models
• Access to large volumes of data for model training
• Applications that leverage the infrastructure to develop real-world use cases

In addition to these technical components, ALIA is built around three guiding principles: it is public, open, and free of profit motives.

A Multilingual and Open-Source Model

ALIA kit provides access to different AI resources and services in Basque (Euskera) and Galician, such as language models, multimodal models (text, voice, image, Spanish, Catalan, Valencian), methodologies, documentation, and training and evaluation datasets. Its open-source nature allows public bodies, businesses, and researchers to build new applications using its underlying technology. By prioritizing linguistic diversity and public accessibility, ALIA marks an important step in Spain’s and Europe’s AI journey.

One of ALIA’s defining features is its open-source model, designed to benefit Spanish-speaking countries beyond Spain. “The goal is not to be the best-performing LLM in the world, but the most widely adopted in the Spanish-speaking world,” said Martorell. A key factor in achieving this is the open-source approach, which brings several advantages. By being openly accessible, both public and private actors can freely access and build on top of ALIA’s infrastructure. The collaborative nature of open-source development also encourages contributions from a wide range of experts, fostering innovation and customization. Additionally, this model allows developers to tailor ALIA to regional contexts, improving its performance in diverse Spanish-speaking communities. ALIA’s open approach is already driving adoption across the Spanish-speaking world, enabling new use cases that address local needs.

What Issues Is ALIA Aiming to Address?

ALIA was developed to address several key challenges:

• AI language gap: Many large language models (LLMs) are trained using information available on the internet, which is predominantly in English. One of the largest open repositories of web crawl data, CommonCrawl, holds less than 5% of its content in Spanish and significantly less in other languages such as Catalan, Valencian, Basque (Euskera), or Galician.
• Fostering innovation: ALIA seeks to facilitate innovation via an ecosystem of AI startups that integrate ALIA models to provide accessible, locally adapted, and AI Act-compliant AI solutions.
• Reduce technological dependence: By building sovereign AI infrastructure, Spain aims to reduce its reliance on foreign technologies and global supply chains. ALIA infrastructure is aligned with the European strategy to promote innovation and the adoption of technologies, guaranteeing technological reliability and social and economic inclusion.
• Enhance public service delivery: ALIA supports the development of transparent and responsible AI tools that can improve service delivery and benefit citizens in areas such as health, public finances, or transport.

Overcoming Challenges in ALIA’s Development

Developing and operating ALIA has also involved navigating obstacles inherent to a project of this nature. Martorell highlighted four main obstacles:

• Data: Building the necessary data resources took five years. This involved creating a hardware platform to manage large-scale data downloads and establishing partnerships with institutions such as libraries and public administrations to access the data needed to train the models.
• Computing Power: Developing ALIA required access to a powerful supercomputer. Without a large GPU-based infrastructure to handle simultaneous data processing, training a model of this scale would not have been possible. Luckily, the BSC is home to MareNostrum 5, Spain’s most powerful supercomputer. Still, managing computing resources proved complex, with unexpected interruptions and unpredictable demand posing additional hurdles.
• Talent: The project relied on a team of only 50 experts in language technologies. Attracting, retaining, and training this talent proved to be a significant challenge, especially in a tech-dynamic ecosystem where publicly funded projects have to compete with big AI labs and companies for talent.
• Political Support: Securing long-term government commitment was critical. Convincing decision-makers of the need for a public AI infrastructure—and the sustained investment required to maintain it—was no small task.

What’s Next?

Looking ahead, Martorell foresees three main categories of use cases that could emerge from ALIA. First, this infrastructure will be particularly valuable for developing AI systems that support medical diagnosis, improving public health delivery. Second, ALIA can facilitate the integration of chatbots into public platforms—a capability that is widely used in the private sector but still emerging in public services. Lastly, ALIA will enable public administrations to better leverage their existing datasets. Governments often maintain extensive information banks that remain underutilized due to a lack of tools to process them efficiently. By addressing these gaps, ALIA has the potential to unlock new opportunities for innovation in public services.

ALIA will also be particularly useful for small and medium-sized companies, enabling them to develop new applications, optimize their processes, or create new products or services that can compete in international markets. ALIA will also be of particular relevance for the recently announced BSC AI Factory, a core pillar for the European Commission to accelerate the development of AI applications aligned with European values. In the short term, ALIA is also expected to support governments in Latin America to develop AI systems aligned with their context.

Ensuring the responsible use of ALIA will be equally important. Given the challenges of enforcing oversight and accountability for open-source models, the BSC is taking proactive steps to mitigate risks. The BSC works closely with the oversight body, the Spanish Agency for AI Oversight (AESIA), to promote safe and ethical development. Additionally, ALIA is designed to align with the EU AI Act’s requirements, reinforcing its commitment to responsible AI practices.

ALIA demonstrates how investment in public AI infrastructure can support technological innovation while promoting cultural and linguistic diversity. By adopting an open-source approach and prioritizing accessibility, ALIA has the potential to become a valuable tool not only for Spain but also for Spanish-speaking communities worldwide.