Bielik 11B

Bielik 11B is an open-weight large language model built for Polish and other European languages, developed by the SpeakLeash Foundation together with the Academic Computer Centre Cyfronet at AGH University of Krakow. Training runs on Poland's Helios and Athena supercomputers, and the v3 generation introduced a dedicated Polish-optimized tokenizer with FOCUS-based embedding initialization, replacing the earlier universal Mistral tokenizer to cut inference cost and extend effective context for Polish text.

The post-training pipeline uses supervised fine-tuning, direct preference optimization, and group relative policy optimization with verifiable rewards, with emphasis on STEM and mathematical reasoning. Released under Apache 2.0 with the v3.0 instruct variant available in 2026, it ships through Hugging Face, Ollama, and AWS SageMaker JumpStart. The team also publishes the SpeakLeash Polish text corpus, the largest documented Polish dataset.