Case Study on ALIA

This case study has been drafted based on comprehensive desk research—including official documentation, technical portals, and academic articles—as well as an interview conducted by the OSOR Team with the Spanish State Secretariat for Digitalisation and Artificial Intelligence (SEDIA), part of the Ministry of Digital Transformation and Civil Service. 

In line with the G7 Vision on AI Openness, ALIA can be understood as aligning most closely with the “Open Source AI” tier of the AI openness spectrum: its model weights, deployment code, and training code are made available under the Apache 2.0 licence, while its training data is governed by a mix of licences depending on the source. This distinction is worth keeping in mind when referring to ALIA as “open” or “open source” throughout the case study.

As of 2026, the AI landscape is predominantly monolingual. Approximately 90% of the training tokens used to build popular large language models (LLMs) are extracted from English texts. [1] This creates gaps in the performance of many existing AI models in other languages and inadvertently marginalises non-English speakers, limiting the utility of these digital tools in non-English or multilingual environments like the European Union (EU).

Beyond the linguistic dimension, European actors also face challenges linked to reliance on proprietary AI models, most of which are developed outside Europe. Currently, closed source solutions account for up to 80% of model token usage, despite the fact that research shows open source alternatives achieve 90% of their performance at the time of release, and rapidly close the remaining gap. [2] Indeed, while open models often have lower parameter counts than large proprietary systems, techniques such as Retrieval-Augmented Generation (RAG) and fine-tuning can reduce performance gaps by grounding model outputs in relevant data or adapting models to particular tasks. In addition, the dependency on proprietary models represents a significant financial burden, as closed source models are, on average, six times more expensive than their open counterparts. [3]

In this technical context and driven by a determination to provide a solution to the linguistic challenge—which is acutely felt in Spain, where four co-official languages coexist with Spanish—the Spanish Government has taken decisive action through ALIA, a public AI infrastructure developed as part of the National Artificial Intelligence Strategy (ENIA) 2024. By providing Europe’s first publicly developed, open and multilingual AI infrastructure, ALIA supports linguistic diversity in LLM development, reuse, and deployment in Spain and across Europe. The initiative aligns with the EU Digital Decade 2030 targets and with the broader direction of the recently released European Tech Sovereignty Package, which emphasises strengthening Europe’s capacity to develop, deploy and maintain digital technologies across different layers of the digital stack, including AI, and positions open source as a key enabler of transparent and sovereign digital ecosystems. 

ALIA is a comprehensive, publicly funded AI infrastructure that serves as a foundational digital public good and represents the materialisation of Spain’s National Artificial Intelligence Strategy. ALIA’s stated purpose is twofold: (1) to provide a public infrastructure of AI resources, such as open LLMs, and (2) to foster the advancement of Spanish and Spain’s co-official languages in the development and deployment of AI worldwide. 

At its core, the ALIA project revolves around the creation of open source foundational and multimodal AI models, encompassing text, voice, and image capabilities. The models are supported by the ALIA Kit, a dedicated digital space that centralises access to technical documentation, demonstrators, training datasets and integration materials. The ALIA Kit acts as the access layer, helping developers, users and administrators understand, adapt, use and reuse the resources. 

In practice, ALIA allows for full rights of access to the core model weights under the Apache 2.0 licence, alongside publicly available training scripts and configuration files under the same terms. This allows for the use, modification, and redistribution of the ALIA models by administrations, companies, universities, and civil society alike. However, while the structural components of the models are uniformly open, the training data itself is governed by a patchwork of licences depending on the source. This includes public domain datasets like CC0 Common Crawl, GitHub repositories under their original licences, Wikimedia under CC BY-SA 4.0, EurLex under CC BY 4.0, and various other corpora obtained through specific permissions or legal agreements. This shapes how ALIA can be reused in practice: the models and code provide a permissive basis for adaptation and deployment, while the underlying datasets require a case-by-case assessment.

The ALIA project focuses on developing and making available highly advanced LLMs—AI systems specifically designed to comprehend, generate, and interact with human language. The best-known model in the ALIA ecosystem is ALIA-40B, a mid-size, [4] multilingual foundational AI model featuring 40 billion parameters that supports a broad range of linguistic tasks. The Barcelona Supercomputing Centre (BSC) describes ALIA-40B as one of Europe’s most sophisticated public multilingual foundational models to date. [5] Designed to support and reflect the rich linguistic diversity of the EU, ALIA-40B has been pre-trained on a large, highly curated corpus of 9.37 trillion tokens encompassing 35 European languages and 92 programming languages. [6] This design confirms ALIA’s spirit of a Spanish-led, but not Spain-exclusive solution. 

Pre-training corpus composition of ALIA-40B. Source: https://huggingface.co/BSC-LT/ALIA-40b 

Alongside ALIA-40B, the ALIA ecosystem includes the Salamandra model family, optimised for practical, agile deployment. Featuring smaller versions with lower infrastructure requirements than ALIA-40B, Salamandra is highly accessible for real-world applications, including translation tasks through SalamandraTA or ensuring safer multilingual deployments across public service applications through Salamandra-Guard.

Finally, the ALIA project comprises the RoBERTa model family. Based on the RoBERTa architecture, these models provide an additional set of multilingual resources for language-understanding and evaluation-related tasks. Their development forms part of a broader line of work that also includes the generation and curation of training corpora and a model evaluation framework, reinforcing ALIA’s role not only as a provider of AI models, but also as an infrastructure for testing, comparing and adapting multilingual AI resources.

Further details on these models and model families can be found in the OSOR article titled Spanish Government promotes open access to its ALIA models.

The successful development of the ALIA models relies on two technical foundations: comprehensive multilingual datasets and high-performance computational power. The project utilises large textual corpora for training, including the CATalog dataset, which contains approximately 23 billion tokens, as well as specialised bilingual administrative data like the ALIA_DOGV corpus. Processing this volume of data is made possible by the MareNostrum 5 supercomputer at the BSC, which supplies the computational capacity required for large-scale AI training. 

Picture of the MareNostrum 5 Supercomputer at the Barcelona Supercomputing Centre. Source: https://www.bsc.es/marenostrum/marenostrum-5

Beyond data and computing infrastructure, the ecosystem’s practical utility hinges on the functionalities that make its models usable in real-world settings. First, agentic integration allows ALIA to transcend simple text generation: through function-calling features, it directly connects the models with external APIs, databases, and administrative workflows to power functional assistants. Second, instruction-following adaptation ensures daily usability by training the models to respond reliably to specialised public-service tasks, standardised forms, and citizen inquiries.  

According to SEDIA, practical use cases and deployments of ALIA are actively emerging across multiple tasks and fields. A primary driver of this is the GobTechLab initiative, which is currently identifying and testing approximately 19 high-impact AI use cases designated specifically for the Spanish public sector. This effort helps public administrations evaluate how generative AI can best optimise internal workflows, enhance service delivery, and improve direct citizen interactions. On the local level, the infrastructure powers generative AI assistants for municipalities, enabling them to help citizens navigate complex procedures, answer routine inquiries, streamline document drafting, and translate bureaucratic jargon into clearer language.

Two flagship initiatives demonstrate the ALIA ecosystem’s versatility in critical public fields:

• Taxation and customs: through the Hacienda initiative, advanced tools built on ALIA are being explored to optimise internal administrative workflows and elevate the quality of assistance provided directly to taxpayers.

• Healthcare: the Cardiomentor initiative adapts ALIA’s capabilities to assist primary care professionals in the medical field, providing them with intelligent support for the early detection of heart failure. 

ALIA operates through a collaborative model that brings together public institutions, research centres, and technical experts. 

Institutional governance and strategic direction are driven by the Spanish Government through the Ministry for Digital Transformation and Civil Service, operating through the State Secretariat for Digitalisation and Artificial Intelligence (SEDIA). SEDIA provides financial oversight and the public policy framework of the initiative, supporting its objective of generating sovereign, open and reusable AI assets. To ensure responsible deployment, the Spanish Agency for the Supervision of Artificial Intelligence (AESIA)—Europe’s first dedicated national AI supervisor [7]—ensures that ALIA aligns with emerging regulatory frameworks and ethical guidelines and complies with the EU AI Act.

The technical execution of the project is coordinated by the Barcelona Supercomputing Centre through a network of researchers and language technology experts. In addition to providing access to high-performance computing infrastructure, BSC contributes deep research expertise and manages key technical processes, including the development, maintenance, documentation, and publication of the models, tools, and related resources.

The project is financed through a publicly funded model combining national budget allocations and European funding instruments, notably from the European Regional Development Fund (ERDF). The development phase received approximately €10 million in funding, encompassing the strategic lines of action outlined in Lever 3 of ENIA. The budget supports the establishment of the public AI infrastructure, including model development and training, and allows the Spanish Government and the BSC to ensure the long-term operation, maintenance, and evolution of the ALIA ecosystem. 

The present phase of the ALIA project runs until the end of June 2026, while the next phase is currently being defined. Moving forward, the project is expected to continue focusing on the development of open, sovereign, and reusable AI assets, with the ultimate goal of enabling an AI infrastructure that strengthens the wider European AI ecosystem. This includes the introduction of new models, improved tools, better datasets, and enhanced evaluation capabilities, ultimately delivering effective and efficient services for European public administrations and other users.

According to SEDIA, ALIA’s primary value lies exactly in its contribution to a shared European AI infrastructure. By delivering a comprehensive ecosystem that includes curated data, computing capacity, open models, evaluation frameworks and specialised expertise, the initiative provides a reusable foundation on which public administrations, companies—including SMEs and start-ups—and researchers can develop sovereign AI-based services. In doing so, ALIA may help reduce reliance on proprietary and non-European alternatives and promote AI assets designed around transparency, public oversight, and a deep integration of regional linguistic and cultural heritage. 

At the same time, ALIA faces several challenges:

• Managing expectations: ALIA’s current computational scale and general-purpose capabilities do not yet match those of the largest commercial frontier models. Clear communication is therefore needed on the types of use cases for which the infrastructure is currently most suitable.

• Data governance: managing the full data lifecycle remains complex, particularly when balancing data quality, licensing conditions and representativeness across different languages and domains.

• Public trust and communication: Public engagement is essential to ensure that citizens clearly understand why public AI infrastructure is being developed, how it is governed, and what safeguards are in place to support safe and responsible use.

These challenges are not specific to ALIA. They reflect broader issues faced by public and open AI initiatives, especially those seeking to combine multilingual coverage, openness, public oversight and practical deployment. The project has the institutional backing, technical expertise, and political momentum for addressing these issues over time. ALIA’s future impact will then depend on how effectively the project continues to improve its models, datasets, evaluation methods and adoption pathways.

For information regarding the policy context of open source software in the Spanish public sector, please consult the respective Country Intelligence Report and its corresponding factsheet in the OSOR Knowledge Centre. These resources include a detailed overview of the political actors, strategic players, political and legislative frameworks and public sector open source software initiatives in Spain. 

[1] Li, Z., Shi, Y., Liu, Z., Yang, F., Liu, N., & Du, M. (2024). Quantifying multilingual performance of large language models across languages. arXiv preprint arXiv:2404.11553. 

[2] [3] Eastwood, B. (2026, January 20). AI open models have benefits. So why aren’t they more widely used? MIT Sloan School of Management. https://mitsloan.mit.edu/ideas-made-to-matter/ai-open-models-have-benefits-so-why-arent-they-more-widely-used  

[4] ApX Machine Learning. (n.d.). Model size examples. https://apxml.com/courses/llm-model-sizes-hardware/chapter-1-intro-llms-model-size/model-size-examples

[5] Barcelona Supercomputing Center. (2025). ALIA, Europe’s first public, open and multilingual AI infrastructure. https://www.bsc.es/news/bsc-news/alia-europes-first-public-open-and-multilingual-ai-infrastructure.

[6] Language Technologies Lab from Barcelona Supercomputing Center. (2025). ALIA-40B Model Card. https://huggingface.co/BSC-LT/ALIA-40b.

[7] Computerworld España. (2025, August 1). Ignasi Belda (AESIA): España tiene un potencial enorme en IA, lo único que nos hace falta es creérnoslo. https://www.computerworld.es/article/4032845/ignasi-belda-aesia-espana-tiene-un-potencial-enorme-en-ia-lo-unico-que-nos-hace-falta-es-creernoslo.html