The race to build genuinely multilingual AI is uneven. Frontier models that handle English, French and Spanish with near-perfect fluency still falter in Finnish, produce off-register Hungarian, and struggle to distinguish Norwegian Bokmål from Nynorsk. The reasons are structural: less training data, more demanding morphology, and a quality ceiling that simply scraping more of the web cannot raise.
Letrum Linguistics, founded in 2026 in Cesson-Sévigné, France, was built to address exactly this problem. The company is the specialist AI training data provider for seven Nordic and Central European languages: Swedish, Polish, Hungarian, Czech, Danish, Norwegian and Finnish. It does not work in other languages. That focus is the foundation of its value proposition.
A category that needs a specialist
For the major languages of the web, large-scale crowdsourcing has been the dominant model for AI data production. It works because the available pool of annotators is deep enough to absorb individual error. For the seven languages Letrum Linguistics serves, the math falls apart. The total native-speaker population across all seven is around 75 million, with Polish accounting for more than half. Skilled annotators are scarce, and the demands placed on them are higher: Finnish and Hungarian agglutination, Slavic verbal aspect, Norwegian biscriptality, Danish phonology that breaks speech systems trained on text. These are not problems a generalist platform can solve with more headcount.
Letrum Linguistics’ contributors are not crowdworkers. They are native-speaking linguists and professional translators, qualified through a multi-stage process inherited from the company’s parent group, Ellipse World, a professional translation operator active since 2017 in France, Canada and the United Kingdom. The vetting standards, project management workflows and quality assurance routines are transposed directly from professional translation, where the cost of error is high and accountability is contractual.
Three product lines, one focus
Letrum Linguistics structures its offer around three product lines, each addressing a different stage of AI model development.
Training Corpora
Training corpora cover original written content, produced or curated to client specification, with provenance documented at the document level. This is the foundational layer: the raw material on which a model’s competence in a given language is built.
Evaluation Sets
Evaluation sets are designed to test specific capabilities of a model, from general fluency and factual accuracy to narrower domains such as legal terminology, medical register or technical writing. They are constructed in collaboration with the client’s own evaluation team, which means the metrics that matter to the lab are the metrics the dataset is built to measure.
Preference Data
Preference data is the most strategically important of the three. It powers reinforcement learning from human feedback and the alignment work that distinguishes a competent model from one that is genuinely usable. It is also the segment where the gap between crowdsourced and expert annotation is widest, because preference annotation requires the kind of judgment that takes years to develop and cannot be reduced to a rubric.
Built for the labs
The expectations Letrum Linguistics is built to meet are those of frontier AI labs and large enterprise buyers. The company operates under European data residency by default, signs mutual non-disclosure agreements at the evaluation stage, and provides a GDPR Article 28 Data Processing Addendum on request. The contractual framework is designed to clear procurement without negotiation, which matters more than it sounds: in this market, the gap between first contact and first project is often measured in legal review cycles, and shortening that gap is a competitive advantage in itself.
Norwegian is operated as two separate pools, one for Bokmål and one for Nynorsk, in recognition of the legal and educational status of both written standards. Generalist providers routinely flatten this distinction. Treating it correctly is a small detail that signals everything about the company’s posture.
Why specialization wins here
The argument for a generalist data provider is scale. The argument for a specialist provider, in a category like this one, is that scale is the wrong axis. The seven languages Letrum Linguistics serves are not a long tail. They are seven national languages of developed European economies, several of them inside the European Union, all of them under regulatory pressure to be served properly by AI systems deployed on their territory. They are also, collectively, a market large enough to sustain a dedicated company built around them.
For frontier labs, the operational question is whether to absorb the cost of building this competence internally or to source it from a provider that has already done so. For enterprise buyers deploying AI products in Nordic and Central European markets, the question is whether to accept that their users will encounter subtle linguistic failures or to invest in the data work that prevents them.
Letrum Linguistics is positioned as the answer to both questions. Its specialization is not a marketing posture but a constraint that runs through hiring, workflow design, quality control and pricing. Every contributor, every process and every contract is built around the same seven languages. That coherence is what a generalist cannot replicate at any price.
Looking ahead
The company is at the start of its commercial trajectory, but the foundation is unusually solid. The infrastructure, the contributor base and the operational standards are inherited from a translation group with eight years of multi-jurisdiction experience, which removes most of the ramp-up risk that affects new entrants in this space. The bet Letrum Linguistics is making is that the AI data market, having matured past the phase of indiscriminate scaling, is now ready to pay for specialization where specialization actually matters.
For the seven Nordic and Central European languages it serves, the case is already self-evident.
