Bielik 11B v3

Bielik is Poland's leading open-source large language model, built by the SpeakLeash community with supercomputing support from Cyfronet AGH in Kraków. The v3 generation (Bielik-11B-v3.0-Instruct, released end of 2025) introduced the family's first dedicated Polish-optimized tokenizer, replacing the previous Mistral-based vocabulary — cutting inference costs and improving effective context for Polish text. Training uses FOCUS-based embedding initialization plus SFT, DPO, and GRPO post-training.

On the Open PL LLM Leaderboard it ranks among the top Polish-language models, reportedly outperforming much larger systems like Llama-3.1-70B-Instruct on Polish linguistic and cultural understanding, while remaining deployable on consumer GPUs via quantization. The 1TB+ Polish corpus was built AI-Act-compliant and released under Apache 2.0.

Honest limits: it is a Polish-first model — English and other-language performance lags generalist models of the same size. Like all open models it can hallucinate, and the ecosystem (tooling, evals) is far smaller than for Llama/Qwen derivatives.