ALIA and AMALIA: Two Open Iberian Models and What a Business Can Do With Them
Spain and Portugal now have public, open language models. What their model cards say, how much memory they need and when they are not worth using.
What they are, and what each says about itself
Portugal and Spain now each have a public, open language model. It helps to know what each one is before deciding whether it fits. ALIA comes from the Barcelona Supercomputing Center, through its language technologies lab. ALIA-40b has about 40 billion parameters (40.43 according to the model card), was pretrained on 35 European languages and then went through post-training concentrated on Spanish, Catalan, Basque, Galician and English. The most recent instruct version we saw is 2606, from June 2026, and quantised GGUF versions exist as well. The licence is Apache 2.0. AMALIA is the Portuguese Government's model, presented in its final version on 1 July 2026. It was trained for European Portuguese, and the Government says it understands text, documents, images and speech. The project site lists education, culture and museums, media and science as areas of application. It is available as open source and, according to the project site, the materials are under the Apache 2.0 licence, with datasets on Hugging Face. On the pages we read we did not find AMALIA's sizes or versions, so we do not quote any.
What the model card says
Model cards are the most honest part of any release, and ALIA's says more than the press does. It says, to begin with, what the model was not designed for: autonomous function calling, specialised code generation and advanced reasoning. For a company that wants to build agents that call tools, this is first-order information and it sits on one line of the card. It also says that, despite safety and value alignment, the model may occasionally produce unintended outputs, and that the red-teaming assessment had an average attack success rate of 4.5%. It recommends sampling temperatures between 0 and 0.2 and discourages repetition penalties. And it warns, unambiguously, not to deploy it in critical applications without extensive testing and mitigation. None of this is a hidden flaw: it is a document doing its job. It is worth reading the model card of whatever you plan to use before reading any news about it, this article included.
The hardware arithmetic
The question that decides a lot is the dullest one: does it fit on the server? We did the sum, and it is arithmetic, not a test. A model keeps every parameter in memory. With 40.43 billion parameters, at 16 bits (2 bytes each) the weights alone take close to 81 GB. At 8 bits, about 40 GB. At 4 bits, roughly 20 GB. And that is before counting the memory of the conversation: the model supports contexts up to 163,840 tokens, and long contexts multiply consumption. The model documentation recommends vLLM for full-precision inference and llama.cpp for the quantised versions. They are the right tools, but the numbers above rule. An entry-level VPS like the one in our Llama 3 8B guide will not run this model. It needs a GPU with a lot of memory, or several, and that changes the cost by an order of magnitude. The sums are ours and they are a lower bound: weights only. We measured nothing on this model.
| Precision | Bytes per parameter | Memory for weights only |
|---|---|---|
| 16 bits (half precision) | 2 | ~81 GB |
| 8 bits (quantised) | 1 | ~40 GB |
| 4 bits (quantised) | 0,5 | ~20 GB |
When it makes sense
There are three reasons that survive analysis, and a fourth that is usually just enthusiasm. The first is data sovereignty. A model running on your own infrastructure does not send customer text to a third party, and that simplifies the GDPR conversation about processors. It does not remove it, but it shortens it. The second is language. ALIA covers Galician, Catalan and Basque in post-training, which is rare. AMALIA was built for European Portuguese, and anyone who has watched a general-purpose model answer in Brazilian Portuguese knows the value of a defined target. That said, we have no numbers showing one is better than another in a specific case, and we are not going to invent them. The third is cost predictability at high volumes: once the hardware is paid for, each extra request costs little. It only pays off above a certain volume, which depends on your case. The fourth, the one we distrust, is: it is public, so it must be better. Public and open means you can use it, inspect it and host it. It does not mean it answers your questions better than the model you already use.
When we say no
There are also cases where we tell a client no. If the job is hard reasoning, code generation or agents that call tools autonomously, ALIA's own card says it was not built for that. A frontier model over an API will do it better, and the cost difference rarely justifies the quality difference. If the volume is low, an API is almost always cheaper than a GPU rented 24 hours a day to answer a few dozen requests. If nobody in the company knows how to run a GPU server, the real cost includes the time of whoever will maintain it, and that time rarely makes it onto the spreadsheet. And if the application is critical, ALIA's card asks for extensive testing before it goes into production. That holds for any model, but here it is written in the official document.
Openness does not switch the law off
Using an open, public model removes no obligation from anyone. A chatbot built on ALIA or on AMALIA is still an AI system that talks to people, and Article 50 of the AI Act applies to it exactly as to one built on any other: the first message has to say it is artificial intelligence. We covered that in the article on the Digital Omnibus. The Apache 2.0 licence permits commercial use, and that is what it does. It does not transfer responsibility for what the model says: that stays with whoever puts it into production. And GDPR still applies. Hosting the model in-house changes who is the processor, not the need for a legal basis, a record of processing and appropriate technical measures for the data passed to it.
How to test before deciding
If you want to know whether one of these models fits, the test is cheap and does not need a server of your own to start with. Gather fifty real cases: questions your customers ask, documents your staff summarise, texts you have to draft. Fifty is enough to see patterns and too few to fool yourself with statistics. Run the same fifty through the model you use today and through the candidate, without telling whoever scores them which is which. Count only the errors that matter to your business: a wrong figure, a promise the company does not make, an answer in the wrong language. Ignore style. Measure latency and work out the cost per thousand requests for both options, hardware included. Then decide with what you measured. If the difference does not justify the change, staying where you are is a perfectly good answer.
Pin the version and question the supplier
Something ALIA's Hugging Face listing shows that tends to go unnoticed: several versions are in circulation. We saw the instruct versions 2512, 2601 and 2606, plus a function-calling variant also from June 2026. The numbers are release months, and each version may answer the same question differently. The practical consequence is to pin the version. If a production system points at the latest one, its behaviour can change without anybody having decided that, and the fifty-case test stops being valid the day the model is replaced. Record which version you tested, and repeat the test before changing it. If whoever proposes a chatbot on one of these models is a supplier, there are four questions worth asking in writing. Which model and which version, exactly. Where it runs and in which country the data ends up. What happens to the conversation logs and for how long. And how and when they will retest when the model changes. A vague answer to any of them tells you more about the supplier than any demonstration.
Frequently Asked Questions
Are ALIA and AMALIA free for commercial use?
According to the sources we consulted, both are under the Apache 2.0 licence, which permits commercial use. The cost is in the infrastructure to run them, not in the licence. Read the licence in the repository of the model you choose before building on it.
Can I run ALIA-40b on a cheap VPS?
By the arithmetic, no. The weights alone take about 81 GB at 16 bits and roughly 20 GB at 4 bits, before counting context memory. It is our own lower-bound estimate, not a measurement.
Does AMALIA suit other Portuguese variants?
It was trained for European Portuguese. We have no data on its performance in other variants and are not going to assume any.
Which is better, ALIA or AMALIA?
There is no direct comparison here: they serve different languages and we have tested neither. It depends on the language and the use case, and the only way to know is the test with real cases described above.
Does hosting the model in-house exempt me from GDPR?
No. It reduces the number of processors to declare, but a legal basis, a record of processing and appropriate technical measures are still needed. This article is not legal advice.