/08 · AI

Bielik AI

A national language model is infrastructure, not software. It requires a corpus, a machine and a jurisdiction — and states have begun to treat all three as sovereign property.

Fig. — Bielik AI Pignus

Sovereign compute · national-language pretraining

/01

The frontier

What is actually happening at the edge.

/01

The corpus is the constraint

Pretraining a competent national-language model is a data problem before it is a compute problem. Open web crawls yield Polish in a fraction of the volume they yield English, much of it duplicated or machine-translated, so the work sits in curation: deduplication at scale, quality classification, OCR of archives and legal corpora, and negotiated access to press and publishing back-catalogues. Tokenisers fitted principally on English fragment Polish inflection and inflate tokens per word, taxing training and inference alike; refitting or extending the vocabulary is now standard practice, and forces the second decision — continued pretraining from an English-centric base, or from scratch. Synthetic and translation-augmented data close part of the gap at the cost of translationese and a narrowing distribution.

/02

Compute has a nationality

Europe's answer has been public rather than commercial: the EuroHPC systems — LUMI, Leonardo, MareNostrum 5, the exascale JUPITER — extended by the AI Factories programme, with national academic machines carrying much of the national-model work, Poland's among them. These are fair-share batch schedulers being asked to perform industrial training, where month-long checkpoint-heavy runs make node failure, storage bandwidth and interconnect the binding limits rather than peak FLOP. Accelerator supply remains subject to export control and allocation politics, which makes secured capacity, not model architecture, the scarce input.

/03

Open weights as governance

The split is no longer open against closed on principle, but on what an institution must be able to verify. Open weights let a ministry, a bank or a hospital run a model inside its own perimeter and inspect it; they do not by themselves confer reproducibility, which additionally requires the data mixture and the training recipe. Licences run from Apache-2.0 to bespoke community terms carrying use restrictions, and the EU AI Act's general-purpose obligations — a summary of training content, systemic-risk duties above the 10^25 FLOP threshold, with relief for genuinely open releases that does not extend above that threshold — have made the licence a strategic instrument rather than a legal afterthought.

/04

Proof in the language itself

Benchmarks translated from English measure translation as much as capability, and contamination is endemic; credible assessment now rests on natively authored tests — state examinations, professional licensing papers, legal and administrative corpora — and on blind human preference at scale. For public administration the bar is different again: answers grounded in cited acts and rulings, refusals calibrated rather than reflexive, and an audit trail that survives appeal. Serving decides the remainder, where quantisation, paged attention, speculative decoding and batching economics determine whether a regulated body can afford the model it has chosen.

/02

Our position

How the families stand in this field.

The vertical is approached as infrastructure, not software. A national model rests on four layers — corpus, compute, power and deployment — and origins in railways, energy and fuel make the middle two legible in a way that software capital rarely finds them. Positions are taken across the stack, on the assumption that the layer which proves scarce will not be the one currently in fashion.

Sovereignty is the asset. A model trained on a nation's own record, held under its own law and served inside its own borders behaves less like an application than like a right of way: bound to a jurisdiction, difficult to substitute, and valuable for as long as that jurisdiction exists. The same requirement is surfacing in every jurisdiction where the house is present.

The horizon is measured in model generations, not releases. Capital that reports quarterly cannot fund a training run that fails, a corpus that takes years to clear, or an institution that becomes indispensable only in its second decade. In a period when states, currencies and institutions are all in motion, that patience is the position.

/03

Other briefs

Participation in the fund is by introduction, from USD 100 million, for a minimum of three years.

Request an introduction