Albanian deserves frontier AI

When people talk about bringing AI to smaller languages, the conversation often takes on the tone of a charitable afterthought. The serious work happens in English and a few other major languages; the rest of the world is served later, by translation layers and goodwill. We think this framing is wrong, both morally and technically. Making frontier models genuinely good in a language like Albanian is not a courtesy extended once the real problems are solved. It is one of the real problems, and it exposes weaknesses in how models are built, trained and evaluated that the largest languages tend to hide.

Why this is a frontier problem

A model trained overwhelmingly on English can often produce grammatical Albanian. That is a genuine achievement, and it is easy to mistake for competence. But fluency on the surface conceals a great deal. The model may understand a question in Albanian by quietly mapping it onto English concepts, answer from knowledge that is thin or wrong about the region, and render the result in a register that no native speaker would use. It may handle standard written Albanian and fall apart on the Gheg spoken across Kosovo and much of the north. It may cost more to run in Albanian, because its tokenizer breaks Albanian words into more pieces, and it may therefore see less of a long document before its context fills up.

Each of these is a research question in its own right. How much can a model learn about a language from limited data, and how much can it borrow from others? How do you measure quality when there are few trustworthy benchmarks and translated ones mislead? How do you represent dialectal variation without flattening it into a single standard? How do you design tokenizers and training mixtures that do not penalise languages for being smaller? These are questions about data efficiency, transfer, evaluation and representation, which is to say they are questions about the core of the field. A language that is well served by abundant data lets you avoid them. Albanian does not.

Low-resource languages are where the hard parts of language modelling stop being hidden.

What happens when the best tools do not speak your language

Tools shape what people do with a language. When the best writing assistants, search engines, tutoring systems and voice interfaces work beautifully in English and adequately in Albanian, people draw the obvious conclusion. Students write in English because the feedback is better. Professionals draft in English and translate afterwards. Developers build products in English first because the models behave more predictably. None of this is dramatic on any given day, but the cumulative effect is to make a language feel less suited to serious work, and that perception becomes self-fulfilling.

The effect compounds through digital presence. Languages that are used less online produce less text, which means less training data for the next generation of models, which means weaker tools, which means less use. Albanian already sits on the wrong side of this loop compared with the major European languages, and it would be easy for the arrival of powerful AI to widen the gap rather than close it. The opposite is also possible. Good tools can make it easier, not harder, to write, learn, publish and work in Albanian, and to do so without switching languages at the first sign of difficulty.

Education, public services and the diaspora

The stakes become concrete quickly. In education, an AI tutor that explains mathematics clearly in English and awkwardly in Albanian teaches a student more than mathematics; it teaches them which language is for thinking. A tutor that understands a student writing in their own regional variety, and responds in clear standard Albanian without correcting them for speaking the way their family speaks, does something much more valuable.

In public services, the requirements are stricter still. A citizen asking about a residence permit, a tax obligation or a health appointment needs an answer that is accurate about the specific institutions of Albania, Kosovo or North Macedonia, not a plausible generalisation drawn from other countries. Errors here are not stylistic. They have consequences for real people, and they fall hardest on those least able to check the answer against another source.

Then there is the diaspora. Albanian is spoken by millions of speakers across Albania, Kosovo, North Macedonia, Montenegro and a large diaspora across Europe, North America and beyond. For many families abroad, the language is carried mostly by grandparents and weakens with each generation. Children may understand Albanian but hesitate to speak it, or speak it but never learn to write it with confidence. Tools that converse naturally in Albanian, correct gently and accept mixed language without complaint could make a real difference to whether the language is passed on. Our own offices span Wyoming, Prishtina and London, so this is not an abstraction for us. It describes our colleagues, their families and our own.

What good looks like

It is worth being precise about the goal, because good enough is a moving target. We think a model serves Albanian well when a native speaker can use it for serious work without mentally translating, without working around it and without noticing that it was built mainly for someone else. Concretely, that means several things.

It means natural, idiomatic Albanian, in the right register, rather than English sentence structure dressed in Albanian words. It means reliable knowledge about Albanian-speaking places, history, institutions, literature and everyday life, and honesty about the limits of that knowledge. It means understanding both Gheg and Tosk, and the informal written language that omits diacritics or mixes in English, German or Italian, while still producing clean standard Albanian when that is what the user needs. It means instruction following and safety behaviour that hold up in Albanian as well as they do in English, rather than degrading when the language changes. And it means comparable cost and context: an Albanian speaker should not pay more, or get less, for asking the same question in their own language.

None of these properties can be taken on trust. Each has to be measured, and measured natively, which is why so much of our work is about evaluation as much as about models.

Our commitments

We are a small, independent lab, and we are not going to claim that we will solve this alone. But we can be clear about how we intend to work.

We build with native speakers. Albanian quality cannot be judged from outside the language, and it cannot be judged by a single speaker standing in for everyone. We work with people from Albania, Kosovo, North Macedonia and the diaspora, and we treat their judgement as the ground truth that our automatic metrics have to answer to, not the other way round.

We respect dialects rather than flatten them. The standard language matters, and a model should write it well. But Gheg is not a mistake, and neither are the regional forms, informal spellings and mixed-language habits of speakers abroad. We intend to measure performance across these varieties separately and to treat regressions on any of them as real regressions.

We share evaluation methods. Much of the difficulty in this area comes from the lack of shared, trustworthy ways to measure progress. As our Albanian evaluation work matures, we intend to publish how we design tasks, recruit and brief raters, and score results, so that others can check our claims and build on the methods. Where we can share data responsibly, we will.

And we will be honest about where things stand. This work is early. We will describe what we have learned when we have learned it, and we will not claim more than the evidence supports. A language deserves tools that tell the truth, and so do the people who speak it.

Albanian has survived empires, borders and long periods in which it was barely written down. It will survive the arrival of AI as well. The question is whether it arrives as a first-class language in the most capable systems of the next decade, or as an afterthought. We think that is worth working on, and we think the work will teach us things about language models that no amount of English could.