Warsaw, January 22, 2026 - The Polish open-source AI community has a new champion with the release of Bielik 11B v3, an 11-billion-parameter multilingual model that significantly outperforms predecessors on Polish and other European languages.

Developed by the SpeakLeash collective, Bielik 11B v3 employs advanced depth up-scaling on a Mistral 7B base, effectively expanding capacity while preserving efficiency. The result is a model that scores 71.83 percent on the PLCC benchmark and 65.93 percent on the Open Polish LLM Leaderboard — gains of several points over prior state-of-the-art Polish models.

Beyond Polish, the model supports 31 additional European languages with strong coherence, making it one of the most capable non-English-centric open models to date. The instruct-tuned variant excels at nuanced tasks requiring cultural context, such as interpreting Polish legal texts or generating idiomatic dialogue.

Krzysztof Wrobel, core contributor to SpeakLeash, explained the motivation: “English dominates LLM training data, leaving smaller languages underserved. Bielik demonstrates that targeted continuation pretraining and careful curation can close the gap without massive compute budgets.”

The model is immediately available in full-precision and quantized formats on Hugging Face and Ollama, lowering hardware requirements for local deployment. Early community testing highlights superior reasoning on Polish STEM questions and reduced hallucination compared to translated outputs from larger English models.

This release strengthens Europe’s push for linguistic sovereignty in AI amid growing regulatory scrutiny of U.S.- and China-dominated foundation models. Bielik joins initiatives like the European AI Alliance and could serve as a base for further national models across the continent.

Credit: SpeakLeash team, led by Krzysztof Wrobel and community contributors. Primary sources: Hugging Face model card, Ollama library, LinkedIn announcement.