Bielik 11B v3.1 ships as a 56-language Polish open model

Poland's open-source Bielik family has a new release: Bielik 11B v3.1, which extends language coverage to 56 languages, according to a report from Pasquale Pillitteri published on November 12.

The model is maintained by the Polish open-source collective SpeakLeash and published on Hugging Face under the Apache 2.0 license. The v3.1 Instruct variant is fine-tuned from the Bielik 11B v3 base, keeping the parameter count at 11 billion, which is small enough to run on a single high-memory GPU or a workstation.

Bielik's pitch has never been raw scale but breadth. Its language tags span Polish, English, German, French, Spanish, Italian, Ukrainian and Czech alongside Chinese, Japanese, Korean, Arabic, Hindi and Turkish, for a total of 56 languages.

Alongside the base weights, the release includes GGUF, AWQ, FP8 and MLX quantizations, covering llama.cpp inference, vLLM serving and Apple-silicon devices. For teams that want to host inference themselves rather than rent it from a cloud provider, those builds often matter more than a leaderboard delta.

Seen more broadly, Bielik is a pragmatic example of Europe's sovereign-model approach: rather than chasing parameter leadership, it concentrates limited compute on multilingual data, local-language quality and permissive licensing. As Brussels pushes AI sovereignty and digital support for regional languages, the practical value of such models is that they can be audited, fine-tuned and deployed privately.

What the release does not yet offer is a head-to-head benchmark against comparable open models, and quality across 56 languages is unlikely to be uniform. The next signals to watch are community evaluations in languages beyond Polish, and whether SpeakLeash extends the line to larger or reasoning-specialized variants.