ALIA & Salamandra: Spain's Open Language Models

While most of the open-weight conversation centres on Mistral, Meta and DeepSeek, Spain has quietly built something different: ALIA, a publicly funded family of language models led by the Barcelona Supercomputing Center (BSC) and trained on Europe's own supercomputers. It is one of the clearest attempts yet to build a sovereign, open and genuinely multilingual alternative to US-led foundation models.

This explainer untangles the three things people keep confusing, the ALIA programme, the Salamandra open-weight models and the flagship ALIA-40B, and explains how they were built, how they compare to other open models, and why a public European LLM matters. It sits in our large language models cluster.

What is ALIA

ALIA is a Spanish public initiative to build open, multilingual foundation models and the infrastructure around them. It is coordinated by the Barcelona Supercomputing Center (BSC), the same institution that runs the MareNostrum supercomputers, and is presented as the first public, open and multilingual AI infrastructure in Europe. The programme is backed by more than €240 million in public funding, which puts it in a different category from most academic model efforts.

The point isn't to chase the absolute frontier on English benchmarks. It's to build models that genuinely speak the languages of Spain and Europe, and to keep the weights, the training recipe and the governance in public hands. ALIA covers Spanish plus the country's co-official languages, Catalan, Valencian, Basque and Galician, alongside dozens of other European languages that big US labs tend to treat as an afterthought.

Practically, ALIA is built to be used, not just published. Because the weights are open and the licence is permissive, Spanish public administrations, universities and small and medium-sized businesses can run the models on their own terms, without sending sensitive data to a foreign API. That accessibility is the whole design intent, and it is what separates ALIA from a research demo.

ALIA vs Salamandra vs ALIA-40B

This is where most coverage gets muddled, so it's worth being precise. The three names refer to a programme, a model family and a single flagship model, respectively.

• ALIA is the programme, the umbrella initiative, funding, infrastructure and governance, coordinated by the BSC. It is not a single model.
• Salamandra is the open-weight model family released by the BSC, currently in two freely downloadable sizes, Salamandra 2B and Salamandra 7B. These are the practical, lightweight workhorses for teams that want something they can run without a data centre.
• ALIA-40B is the flagship foundation model, a 40.4-billion-parameter model that is the most capable output of the programme to date.

A simple way to hold it in your head: ALIA is the lab, Salamandra is the everyday range, and ALIA-40B is the headline release. All of them share the same goal of open, multilingual, sovereign AI, and all the model weights are published openly so anyone can inspect, fine-tune and deploy them.

How ALIA-40B was built (MareNostrum 5)

ALIA-40B is a large undertaking by any measure. With 40.4 billion parameters, it was trained over more than eight months on MareNostrum 5, the BSC's EuroHPC supercomputer in Barcelona and one of the most powerful machines in Europe. Training a model this size on public infrastructure, rather than renting cloud GPUs from a hyperscaler, is itself part of the sovereignty argument.

The training corpus is deliberately broad. ALIA-40B was trained on roughly 6.9 to 9.37 trillion tokens spanning 35 European languages plus 92 programming languages, with strong representation of Spanish and the co-official languages of Spain. That multilingual balance is the model's signature: where many open models are English-first with everything else bolted on, ALIA-40B is built multilingual from the ground up.

• Parameters: 40.4 billion (dense foundation model).
• Training data: ~6.9–9.37 trillion tokens across 35 European languages and 92 programming languages.
• Compute: MareNostrum 5, the BSC's EuroHPC supercomputer, over 8+ months of training.
• Licence: Apache 2.0, fully permissive for commercial and public-sector use.
• Release: published on Hugging Face (BSC-LT/ALIA-40b) in April 2025.

The model below summarises ALIA-40B and Salamandra alongside two reference open models, so the positioning is clear at a glance.

Figures are publisher-stated and current as of June 2026. Note the contrast in governance as much as in size: ALIA and Salamandra are public and Apache-licensed; Mistral is private but permissively licensed; Llama is private with a more restrictive community licence. For European buyers, that governance column often matters as much as raw capability.

Why a public European model matters

The strategic case for ALIA is straightforward: it gives Spain, and Europe, a foundation model that isn't aligned with purely US commercial interests, and whose weights and training process are open to public scrutiny. In French policy circles, Mistral plays a similar role as the home-grown champion; ALIA is the public, sovereign equivalent, owned by a research institution rather than a venture-backed startup.

That distinction has real consequences for who can use it and how:

• Data sovereignty. Public bodies can run ALIA models on European infrastructure without exporting citizen data to foreign APIs, a recurring concern under the EU's AI Act and data-protection rules.
• Linguistic fairness. Catalan, Basque, Galician and dozens of smaller European languages get first-class treatment, not the long-tail neglect they often receive from English-first labs.
• Accessibility. Spanish public administrations and SMEs that could never afford to train a frontier model get a capable, free, fine-tunable base to build on.
• Transparency. Open weights and a documented recipe let researchers audit the model, an advantage closed frontier systems simply cannot offer.

None of this means ALIA out-benchmarks GPT or Claude on English reasoning, and the BSC doesn't claim it does. The value is different: it is the question of who controls the model, in whose languages it is fluent, and on whose hardware it runs. For a growing slice of European public-sector and enterprise demand, those are the decisive questions. This is exactly the kind of debate that fills rooms at AI Summit Europe.

ALIA vs Mistral & other open models

It's tempting to rank ALIA against Mistral, Llama and DeepSeek on a leaderboard, but they're answering slightly different questions. Mistral, Meta and DeepSeek optimise for broad capability and developer adoption; ALIA optimises for European multilingual coverage, public ownership and sovereignty. They overlap on the open-weight, Apache-style licensing that makes self-hosting viable.

• vs Mistral. Mistral is Europe's leading private open-weight lab, with larger Mixture-of-Experts models and a fast commercial cadence. ALIA is the public counterpart, smaller in raw scale but stronger on Spain's co-official languages and built without commercial pressure. They are complements as much as rivals in the European stack.
• vs Llama (Meta). Llama 4 has the biggest open-weight ecosystem and extreme context lengths, but it is US-owned and governed by a more restrictive community licence. ALIA trades ecosystem breadth for European sovereignty and a cleaner Apache 2.0 licence.
• vs DeepSeek. DeepSeek pushes open-weight capability per dollar, but is China-origin and draws governance scrutiny from some European regulators. ALIA's appeal is the opposite, a model whose provenance and funding are fully European and public.

If you want the full landscape of how these models stack up on context window, capability and cost, see our dedicated comparison of the best large language models. For where ALIA fits into the country's wider scene, see AI in Spain and the broader map of the top AI startups in Europe.

How to use ALIA and Salamandra

Because everything is open-weight, getting started is refreshingly direct, no waitlist, no API key, no usage approval.

• Download the weights. ALIA-40B (BSC-LT/ALIA-40b) and the Salamandra 2B/7B models are published on Hugging Face and can be pulled directly into the standard transformers tooling.
• Pick the right size. Use Salamandra 2B or 7B for lightweight, on-prem or edge workloads and quick fine-tuning; reach for ALIA-40B when you need the most capable multilingual base and have the hardware to serve it.
• Fine-tune for your domain. The Apache 2.0 licence permits commercial use, so public administrations, universities and SMEs can adapt the models to legal, administrative or regional-language tasks and deploy them on European infrastructure.
• Build sovereign applications. For organisations bound by data-residency rules, self-hosting ALIA keeps inference and data inside your own environment.

If you're evaluating ALIA as part of a wider model strategy, it pairs naturally with the trade-off thinking in our LLM comparison and the regional context in why Barcelona is a European AI hub. The researchers and institutions behind models like these are exactly who you'll meet among our speakers.

Methodology

How this explainer was compiled

This explainer summarises the ALIA programme and its models from the Barcelona Supercomputing Center's published materials and model cards. Names, parameter counts, language coverage, training details and licensing were cross-checked against public sources before publication, including:

• The official Hugging Face model card for BSC-LT/ALIA-40b and the Salamandra model repositories
• Announcements and documentation from the Barcelona Supercomputing Center (BSC) and the ALIA programme
• Spanish public-data and policy sources, including datos.gob.es, and reporting from Science|Business

As of June 2026. Model versions, training figures and language coverage may evolve as the programme releases new checkpoints. Treat specific numbers here as indicative of the published releases and verify against the BSC and Hugging Face before relying on a precise figure.

ALIA won't dethrone the frontier on English benchmarks, and it isn't trying to. What it offers is rarer: a capable, openly licensed, genuinely multilingual foundation model that is owned by the public, trained on European supercomputers and fluent in the languages of Spain. As the continent debates digital sovereignty, that combination, embodied by ALIA-40B and the Salamandra family, makes it one of the most strategically interesting open models in Europe.

If this interests you

• Compare the field: the best large language models ranks ALIA against the wider open and closed model landscape.
• Go regional: AI in Spain and why Barcelona is a European AI hub set the national and city context.
• Map the ecosystem: the top AI startups in Europe and the AI Summit Europe dossier cover the continent's wider scene.
• Meet the builders: browse the speakers and Mistral AI, Europe's private open-weight counterpart.

Sources

• Hugging Face — official model card, BSC-LT/ALIA-40b (parameters, tokens, languages, Apache 2.0 licence, release).
• Barcelona Supercomputing Center — bsc.es, ALIA programme and Salamandra model documentation, MareNostrum 5.
• datos.gob.es — Spanish open-data portal, ALIA programme and public-funding context.
• Science|Business — reporting on ALIA and European public AI infrastructure.