Polish, measured honestly.

A model that scores well on English MMLU can still fail at a Polish morphology puzzle. Polish has seven cases, free word order, and a century of literature that carries oblique cultural reference. English-first frontier models often translate internally; the loss is invisible on English benchmarks but legible in Polish conversation.

PLCC probes cultural and linguistic competence directly — idioms, registers, references to national authors. CPTU-Bench scores comprehension of nuanced literary and scientific prose. The Open PL LLM Leaderboard is the multi-task catch-all. MT-Bench and EQ-Bench in their Polish variants test conversation and emotional inference respectively.

The two home-grown families — Bielik from SpeakLeash, PLLuM from OPI — are Polish-first by design. They do not always top the boards, but they are the cleanest control for how much of a model's Polish performance is earned by training and how much is a translation trick.

How the Polish register is kept.

Scores are pulled live from the Codesota benchmark database; the page has no hand-written numbers. When the underlying benchmark owners update their tables, our page updates on the next request.

Bielik and PLLuM are the two Polish-first families worth cross-tracking explicitly. The tables in § 03 and § 04 show every version we find, across every benchmark they have been evaluated on. Gaps are gaps — we do not impute.

PLCC and CPTU-Bench use different scoring conventions; comparing raw numbers across benchmarks is meaningless. The SOTA tickers group by benchmark for a reason.