📊 Full opportunity report: ALIA. The Spanish answer. on ThorstenMeyerAI.com — validation score, market gap, and execution plan.

TL;DR

Spain’s ALIA-40B, a €240 million public AI project, has been released as Europe’s largest publicly funded multilingual model. While operationally validated, it shows performance gaps compared to Llama 2, highlighting strategic positioning issues.

Spain has officially released ALIA-40B, a 40-billion-parameter multilingual language model, marking the country’s most ambitious public AI initiative to date. The project, funded with over €240 million in public investment, aims to position Spain as a leader in multilingual AI within Europe, emphasizing Spanish-language coverage and open-source transparency.

The ALIA-40B model, trained from scratch on 9.37 trillion tokens across 35 European languages and 92 programming languages, was announced by the Barcelona Supercomputing Center (BSC-CNS) on April 22, 2025. It is publicly available under the Apache License 2.0 on HuggingFace, and is part of Spain’s broader national AI strategy led by the Secretary of State for Digitalisation and Artificial Intelligence (SEDIA). The project is supported by €90 million for MareNostrum 5 hardware upgrades and €150 million dedicated to ALIA integration into industry, making it the largest publicly funded European national AI effort.

Benchmark results indicate that ALIA-40B’s performance on tasks like XNLI and SQuAD is below that of Llama 2, with accuracy scores of approximately 51.77% and 81.53%, respectively, compared to Llama 2’s 66% and 93-94%. These results confirm the structural capability gap highlighted in prior analysis, emphasizing that ALIA’s strategic positioning focuses on multilingual coverage and adoption rather than raw performance. The project’s leadership emphasizes its goal is to promote widespread use in the Spanish-speaking world, aligning with the ‘Position 3’ strategic profile.

ALIA · The Spanish Answer.
DISPATCH / MAY 2026 ESSAY · EUROPEAN SOVEREIGN LLMs · ALIA · SPANISH ANSWER
▲ Standalone Essay EU Sovereign AI · Tier 2 Expansion · May 2026
Standalone Essay 10 · Spanish National-Continuation Pattern · Position 1 vs Position 3 Interrogation

ALIA.
The Spanish
answer.

€240M+ Spanish public funding · ALIA-40B + Salamandra family · 9.37T tokens · 35 European languages + 92 programming languages · MareNostrum 5 · Apache 2.0 release. The largest publicly funded European national-AI project by cumulative scope — and the empirical test case for the Position 1 vs Position 3 strategic-positioning argument.

This is the tenth standalone essay in the European sovereign-LLM track and the third Tier 2 expansion piece. ALIA is Spain’s institutional answer — the largest EU member state by GDP not yet documented in the track. The project markets itself as Position 1 + Position 2 simultaneously — “Europe’s first public multilingual foundational model.” The benchmark evidence (ALIA-40B 51.77% XNLI_en vs Llama 2 66%) confirms the structural capability gap from Finding 1 of the synthesis essay. The Position 3 framing — Martorell’s “most widely adopted in the Spanish-speaking world” — is operationally honest. €90M MareNostrum 5 upgrade + €150M company integration = €240M+ cumulative scope. Apache 2.0 open-source release + AESIA validation + co-official languages oversampling. Both can be true at once. The Spanish public discourse would benefit from explicit Position 3 strategic positioning.

▲ The structural editorial finding · the Position 1 vs Position 3 interrogation
ALIA is the largest publicly funded European national-AI project by cumulative scope · €240M+ Spanish public investment exceeds Portugal AMÁLIA + Italy Minerva + OpenEuroLLM combined. Benchmark evidence confirms Finding 1’s structural capability gap empirically. Martorell’s Position 3 framing — “most widely adopted in the Spanish-speaking world” — is operationally honest. The Spanish public discourse should explicitly reframe ALIA as Position 3 + Position 4 vertical-specialization.
— standalone essay 10 · the spanish answer · may 2026 · interrogating position 1 vs position 3
€240M+
Cumulative Spanish public funding · €90M MareNostrum 5 upgrade + €150M company integration · 100% publicly funded
Largest national-AI public funding scope in Europe · exceeds Portugal + Italy + OpenEuroLLM combined
9.37T
ALIA-40B training tokens · 35 European languages + 92 programming languages · 8+ months on MareNostrum 5
33 TB training corpus · 4,480 NVIDIA H100 GPUs accelerated partition · BSC-CNS coordination
35 + 4
European languages broad coverage + 4 co-official Spanish languages oversampled by factor of 2
Castilian · Catalan/Valencian · Basque · Galician · plus 30+ other EU languages · Apache 2.0 release
Pos 3
Operationally honest strategic positioning · multilingual specialization with Spanish-language oversampling
Martorell: “the goal is not to be the best-performing LLM in the world, but the most widely adopted in the Spanish-speaking world”
ALIA-40B 40B PARAMETERS · 9.37 TRILLION TOKENS · 35 EUROPEAN LANGUAGES · MARENOSTRUM 5 TRAINING SALAMANDRA-7B 12.875 TRILLION TOKENS FROM SCRATCH · FIRST MARENOSTRUM 5 LLM · BSC-CNS APACHE 2.0 APRIL 22, 2025 HISPANIA 2040 RELEASE · PUBLIC CODE PUBLIC MONEY · AESIA VALIDATED CO-OFFICIAL LANGUAGES CASTILIAN · CATALAN/VALENCIAN · BASQUE · GALICIAN · 2× OVERSAMPLED BENCHMARK GAP 51.77% XNLI_EN VS LLAMA 2 66% · 81.53% SQUAD_EN VS LLAMA 2 93-94% PEDRO SÁNCHEZ LAUNCH ANNOUNCEMENT JAN 21 2025 · €240M+ AI STRATEGY 2024 INVESTMENT
The ALIA model family · five distinct models · April 22, 2025 release

Six models. Apache 2.0.

The ALIA family operates as a tiered model portfolio. ALIA-40B is the flagship at 40 billion parameters; the Salamandra family scales down to 7B, 2B and instruct-tuned variants; mRoBERTa provides the foundational multilingual baseline. All released under Apache License 2.0 on April 22, 2025 at the HispanIA 2040 event — “Public Code, Public Money” approach.

The ALIA model family · all training scripts and configuration files publicly available on GitHub
From the HuggingFace BSC-LT collection and the Salamandra Technical Report (arXiv 2502.08489). The most comprehensive open-source release of any European national-AI project — more accessible than Mistral’s selective open-weights, structurally aligned with Apertus’s full open-source architecture.
ALIA-40BFlagship multilingual
40Bparameters
Transformer-based decoder-only · pre-trained from scratch on 9.37 trillion tokens of highly curated data. 35 European languages + 92 programming languages. 8+ months training on MareNostrum 5.
Flagship
multilingual
Salamandra-7BMid-tier general
7Bparameters
Transformer-based decoder-only · pre-trained from scratch on 12.875 trillion tokens. First LLM trained from scratch on MareNostrum 5’s accelerated partition. 35 European languages + code.
First
MN5 LLM
Salamandra-2BCompact deployment
2Bparameters
Same 12.875 trillion token corpus as Salamandra-7B. Compact deployment for resource-constrained environments — edge inference, embedded systems, mobile applications.
Compact
edge
Salamandra-7B-instructInstruction-tuned
7Binstruct
Instruction-tuned on 276,000 instructions in English, Spanish, and Catalan collected from several open corpora. The primary deployment target for application development.
Deployment
target
Salamandra-2B-instructCompact instruct
2Binstruct
Same 276K instruction corpus applied to Salamandra-2B base. Compact instruction-tuned variant for resource-constrained applications requiring conversational capability.
Compact
instruct
mRoBERTaFoundational baseline
RoBERTaarchitecture
Multilingual foundational model based on the RoBERTa architecture. Pre-trained from scratch using 35 European languages + code. Encoder-only baseline for downstream tasks.
Foundational
encoder
Multilingual coverage · 35 EU languages + 4 co-official Spanish languages
Multilingual AI Translation Mastery: Building Accurate, Culturally Sensitive Language Tools and Global Communication Systems in 2026

Multilingual AI Translation Mastery: Building Accurate, Culturally Sensitive Language Tools and Global Communication Systems in 2026

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Four official. Oversampled by factor of 2.

ALIA’s distinctive multilingual coverage strategy. The four co-official Spanish languages are oversampled by factor of 2 in the training corpus — structurally distinct from Apertus’s broad 1,811-language coverage approach. The strategy targets deep coverage of Spanish co-official languages rather than maximum language breadth.

The four co-official Spanish languages · 2× oversampled in training corpus
Plus 30+ other European languages in the broader 35-language coverage baseline. The training corpus distribution detail Bara surfaced is operationally significant: 16.12% Spanish vs 39.31% English — the multilingual scope dilutes the Spanish-specific specialization.
▲ Castilian Spanish
Español
500+ million native speakers globally. Primary language of Spain and Latin America. Spanish-speaking world adoption strategy target. 16.12% of ALIA-40B training corpus.
▲ Catalan (with Valencian)
Català · Valencià
~10 million speakers · Catalonia, Valencia, Balearic Islands, Andorra. AINA project foundational data. CATalog dataset contribution — largest open Catalan dataset globally.
▲ Basque (Euskera)
Euskera
~750,000 speakers · Basque Country and Navarre. Language isolate (not Indo-European). HiTZ Basque Center for Language Technology (UPV/EHU) coordination. Latxa baseline model.
▲ Galician
Galego
~2.4 million speakers · Galicia and parts of Portugal. CiTIUS + Galician Language Institute (ILG) at University of Santiago de Compostela. Carballo model family.
+ 30 European languages35 total in corpus
Broad 35-language coverage baseline: German · French · Italian · Portuguese · Dutch · Polish · Czech · Hungarian · Greek · Romanian · Bulgarian · Croatian · Slovenian · Slovak · Lithuanian · Latvian · Estonian · Finnish · Swedish · Danish · Norwegian · Maltese · Irish · Albanian · Macedonian · Serbian · Bosnian · Welsh · plus contribution to Community OSCAR (151 languages · 40T words). The structural distinction from Apertus’s 1,811 languages — depth over breadth.
Benchmark evidence · structural capability gap empirically confirmed
Mini AI Voice chatbot, smart Voice Assistant, Multiple AI Models, Emotional Interaction, 100+ Stickers, Suitable for Home and Office use, (Black)

Mini AI Voice chatbot, smart Voice Assistant, Multiple AI Models, Emotional Interaction, 100+ Stickers, Suitable for Home and Office use, (Black)

1. Emotional Interaction: This chatbot can recognise and respond to your emotions, offering a more personalised and human-like…

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

ALIA-40B vs Llama 2. 14-point gap.

The empirical evidence Finding 1 of the synthesis essay needed. ALIA-40B at 40 billion parameters with €240M+ public funding and 8+ months MareNostrum 5 training achieves performance below Llama 2 — a 2023 frontier model released approximately 18 months before ALIA-40B. The capability gap is real and consistent with six of seven prior national-project answers documented in the track.

ALIA-40B vs Llama 2 · benchmark performance comparison
From Bara of Tokiota’s analysis published in Silicon. The empirical capability gap confirms Finding 1 across the European sovereign-AI track — six of seven national-project answers operationally below frontier-class performance.
▲ ALIA-40B
51.77%
XNLI_en Natural Language Inference
▲ Llama 2 (Jul 2023)
66%
Same benchmark · same task
▲ Capability Gap
14.23pp
Below 2023 frontier baseline
▲ ALIA-40B
81.53%
SQuAD_en Question Answering
▲ Llama 2 (Jul 2023)
93-94%
Same benchmark · same task
▲ Capability Gap
11.5pp
Below 2023 frontier baseline
The structural implication: The Position 1 framing — “Europe’s most advanced public multilingual foundational model” — is operationally misleading. ALIA-40B’s benchmark performance does not support the framing. Six of seven prior national-project answers operationally confirm the structural capability gap: AMÁLIA, Minerva, Mistral, Aleph Alpha, Apertus, ALIA. Only OpenEuroLLM’s benchmarks haven’t yet shipped. The Position 3 framing is operationally honest.
“The goal is not to be the best-performing LLM in the world, but the most widely adopted in the Spanish-speaking world.” Josep M. Martorell, BSC Associate Director · Oxford Insights interview · April 2025
Pilot applications · two deployment targets announced HispanIA 2040 event
Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls

Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Two pilots. Public administration deployment.

The operational deployment targets that validate the Position 3 + Position 4 framing. Public administration deployment is the structurally credible Position 3 + Position 4 strategic positioning — captive demand from Spanish public institutions where Spanish-language specialization is operationally distinctive.

Two pilot applications · Tax Agency + primary care medicine
From the Interoperable Europe ALIA release coverage. Both pilots target captive Spanish-language public-administration demand — the operationally credible Position 3 + Position 4 deployment pattern.
▲ Public Administration · Tax
Agencia Tributaria Chatbot
Internal chatbot streamlining work of the Spanish Tax Agency and its citizen service. Spanish-language specialization operationally distinctive · captive demand from public-administration deployment · regulated procurement pattern.
▲ Healthcare · Primary Care
Heart Failure Diagnosis
Primary care medicine application · advanced data analysis facilitating heart failure diagnosis. Regulated healthcare deployment · Spanish-language clinical context · AESIA-validated transparency aligned with EU AI Act.

The work is real across the Spanish ALIA case. €240M+ public funding committed. 40B parameter from-scratch model trained on 9.37 trillion tokens. Salamandra family released under Apache 2.0. AESIA validation aligned with EU AI Act transparency standards. Two pilot applications shipped — Tax Agency chatbot and primary care medicine heart failure diagnosis. The Position 1 framing is operationally misleading. ALIA-40B performance below Llama 2 confirms the structural capability gap. The Position 3 framing is operationally honest — Spanish-speaking world adoption, co-official languages oversampling, public administration deployment. Both can be true at once. The Spanish public discourse would benefit from explicit Position 3 strategic positioning.

— Standalone Essay 10 · The Spanish ALIA answer · interrogating Position 1 vs Position 3 · May 2026
Source dossier · the ALIA operational receipts
Colophon · Standalone Essay 10 · Tier 2 Expansion

Set in Source Serif 4 (display), EB Garamond (essay body), IBM Plex Sans & IBM Plex Mono. Standalone essay register · not part of the security franchise. The Spanish national-continuation pattern interrogation extending the synthesis essay’s Position 1 vs Position 3 strategic-positioning argument with empirical operational analysis. Capital-violet dominant register with all six chromatic registers integrated into the multilingual coverage visualization — Castilian violet · Catalan engineering-blue · Basque terminal-green · Galician window-amber · the broader 35 European languages in synthesis-deep · the Position 1 attempt critique in takeoff-orange. Free to embed with attribution.

thorstenmeyerai.com

Standalone essay 10 · European sovereign AI · The Spanish ALIA answer · May 2026

€240M+ · ALIA-40B · 9.37T TOKENS · 35 LANGUAGES · 4 CO-OFFICIAL · APACHE 2.0 · POSITION 3

Premium Hobbies P-51B & Bf109 G-6 "European Theater 1:72 Plastic Model Airplane Kit Twin Pack 138V

Premium Hobbies P-51B & Bf109 G-6 "European Theater 1:72 Plastic Model Airplane Kit Twin Pack 138V

INCLUDES: 2 (Two) Model Kits

As an affiliate, we earn on qualifying purchases.

As an affiliate, we earn on qualifying purchases.

Strategic Positioning and Operational Validation of ALIA

Spain’s ALIA-40B project demonstrates a significant public investment in AI, emphasizing multilingual capabilities and open-source transparency. Although its benchmark performance lags behind Llama 2, the project’s focus on Spanish-language adoption and co-official languages aligns with its strategic goal of widespread regional influence. This underscores a broader European debate over whether to prioritize performance or strategic linguistic coverage in national AI initiatives, with ALIA exemplifying the latter approach.

Spain’s Public AI Investment and European AI Strategy

Spain’s ALIA project is part of a series of national AI efforts across Europe, which include Portugal’s AMÁLIA, Italy’s Minerva, and France’s Mistral. It is the largest publicly funded effort, with €240 million dedicated to a 40B parameter model, trained from scratch and released under open-source licenses. The project is coordinated by the Barcelona Supercomputing Center and supported by the Spanish government’s AI strategy launched in January 2025. Prior efforts in Europe have varied in scope, funding, and focus, with ALIA representing a strategic shift towards multilingual and open models aimed at regional adoption.

“Our goal is not to produce the most powerful LLM in the world but to foster the most widely adopted in the Spanish-speaking world.”

— Josep M. Martorell, ALIA project lead

Performance Gap and Strategic Implications

While benchmark results confirm a structural performance gap between ALIA-40B and Llama 2, it remains unclear how this will impact long-term adoption and regional influence. Additionally, the strategic choice to prioritize multilingual coverage over raw performance may influence future development trajectories and international perceptions of Spain’s AI capabilities.

Next Steps in ALIA Deployment and European AI Strategy

Further benchmarking and real-world deployment will clarify ALIA’s adoption success. The project team plans to expand multilingual applications and industry integration, while ongoing evaluations will assess whether the strategic focus on coverage over performance sustains regional influence. Additionally, Spain’s AI strategy will likely evolve based on ALIA’s operational outcomes and benchmarking results, shaping future national and European AI initiatives.

Key Questions

What is the main purpose of ALIA-40B?

ALIA-40B aims to promote widespread adoption of multilingual AI in Spain and Europe, focusing on Spanish-language coverage and open-source transparency rather than achieving top performance benchmarks.

How does ALIA compare to other European AI models?

Compared to models like Llama 2, ALIA-40B shows a performance gap but emphasizes multilingual coverage, co-official language support, and regional adoption, aligning with its strategic positioning as a national project.

What are the key benchmarks of ALIA-40B?

On XNLI, ALIA scores approximately 51.77%, and on SQuAD, about 81.53%, both below Llama 2’s scores of 66% and 93-94%, respectively, confirming the structural performance gap.

What does this mean for Spain’s AI ambitions?

It indicates a strategic focus on regional influence and multilingual coverage over performance supremacy, shaping Spain’s role within Europe’s AI landscape.

Source: ThorstenMeyerAI.com

You May Also Like

Anthropic says Trump admin has lifted export controls on Claude Fable 5 and Mythos 5

The Trump administration has removed export restrictions on Anthropic’s AI models Claude Fable 5 and Mythos 5, according to the company. Details remain emerging.

Apple Silicon’s Quiet Memory Advantage

Apple Silicon’s unified memory architecture offers a significant capacity advantage for running large AI models locally, despite lower bandwidth and speed.

The 27% Problem: Why Google Wrote a $750M Check to Catch Anthropic

Google announced a $750 million fund to expand its enterprise AI presence, aiming to challenge Anthropic’s current lead with new platform and distribution strategies.

When AI Builds Itself: Inside Anthropic’s Evidence on Recursive Self-Improvement

Anthropic presents data suggesting AI is increasingly capable of automating research and development, raising prospects for recursive self-improvement.