Bielik 11B v3.0 Instruct is an open-source language model released by the SpeakLeash team, distributed under the apache-2.0 license through the Hugging Face repository. It is positioned as a successor to the Bielik v2 series, and its development explicitly drew on lessons from the Bielik v3 Small models, making it part of an evolving Polish-language model family rather than a standalone release. The 11-billion-parameter dense design requires roughly 22.3 GB of VRAM, which is consistent with running the model in half precision on a single high-end consumer or prosumer GPU, and its openly available weights give developers the freedom to self-host, fine-tune, or audit the model locally.

For alignment and instruction following, the model was trained on a large Polish instruction dataset of more than 20 million examples spanning over 17 billion tokens, combining manually verified instructions with synthetically generated ones to compensate for the scarcity of high-quality Polish supervision. The SpeakLeash team applied a DPO-Positive alignment approach using both generated and human-corrected preference examples scored by a reward metamodel, with a curated pool of more than 114,000 examples targeting different response-style qualities across multiple turns. This training lineage makes the model a strong fit for Polish conversational assistants, content generation, summarization, and enterprise workflows that need a self-hostable, open-weight language model with documented alignment practices.