v3.0 · open source · 🇵🇱 Polish model

Polish, open ecosystem of language models.

We build AI models, tools, and datasets that strengthen Poland's and Europe's technological sovereignty. Bielik runs locally, understands Polish language and context, and its ecosystem is developed by the SpeakLeash Foundation, partners, and the community.

6arXiv publications
93Model variants
4.7 MModel downloads
0.1B–11BSize range
03 · Why Bielik

Polish model. Local deployments. Openness without compromise.

01

Polish language and local context

Training on a high-quality Polish corpus, custom tokenizer, Polish benchmarks — the model understands inflection, cultural context, and specialized vocabulary.

02

Local deployments

Models run on your own hardware, on-premise and in isolated environments. No data leakage to external clouds.

03

Data control

Full sovereignty over prompts, contexts, and results. Audit, governance, and compliance with organizational policies — under your control.

04

Openness and transparency

Model weights, model cards, arXiv publications, benchmarks, and training data — all documented and publicly available.

05

European technological sovereignty

Part of an independent AI infrastructure for Poland and the European Union. Open foundations instead of dependence on single vendors.

06

Ecosystem, not just a model

Research, community, safety, evaluation tools, Bielik Arena, Data Dashboard, deployment partners — complete infrastructure around the core model.

04 · Models

The Bielik model family

From lightweight models running on CPU to advanced variants with instruction tuning and tools — choose the version suited to your needs and hardware.

Model Size Type Formats Status
Bielik 11B v3 Family Flagship family · 11B · instruct / base · agentic since v3.1
11B
Instruct · Base
HF · GGUF · FP8 · AWQ · MLX
Current
Bielik-Minitron 7B v3.0 Compressed from 11B · structured pruning + KD
7B
Instruct · Draft
HF · FP8 · GGUF
Current
Bielik v3 Small 1.5B and 4.5B · local runs, edge, education
1.5B–4.5B
Instruct · Base
HF · GGUF · FP8
Current
Bielik Guard / Sójka Polish-language safety classifiers
0.1B–0.5B
Safety
HF
Safety
All models Hugging Face Collections ↗
05 · For whom

Who is already using Bielik?

Enterprise

Companies

  • RAG on company documents
  • Internal assistants
  • Sensitive data anonymization
  • On-premise deployments
  • PoC in 2–6 weeks
Administration

Public administration

  • Technological sovereignty
  • Government assistants
  • Citizen Bielik
  • Sensitive data locally
  • AI education & competencies
Research

Science & community

  • Open weights & model cards
  • 6 arXiv publications
  • Open PL LLM Leaderboard
  • Bielik Arena
  • Data Dashboard SpeakLeash
07 · Research

We don't just publish models. We show how they're built.

6arXiv publications
11BFlagship model
2024–26Years of research
6Polish benchmarks

Bielik grew out of the SpeakLeash / Spichlerz community — a project aimed at creating, cataloguing, and documenting Polish datasets for large language models. Today we continue this tradition of open science: every model has a model card, technical publication, and benchmark results.

08 · Our partners

Building Polish AI together

Bielik AI is built by public institutions, companies, universities, and individual donors. Every form of support — from computational resources to funding — makes a difference.

Strategic partners
Partners
Supporting partner
10 · FAQ

Frequently asked questions

Didn't find an answer? Write to us.

Getting started

The easiest way is Ollama: "ollama run SpeakLeash/bielik-11b-v3.0-instruct". Alternatively, download the model from Hugging Face and run it via llama.cpp, vLLM, or transformers. Detailed instructions are in the Models section.

License & availability

Yes. Model weights, model cards, evaluation tooling code, and technical publications are publicly available on Hugging Face and GitHub. The Apache 2.0 license allows commercial use without licensing fees.

Yes. The Apache 2.0 license permits commercial use, modifications, and redistribution. There are no licensing fees or deployment restrictions — on-premise, cloud, or edge.

Models

For most production use cases we recommend Bielik 11B v3.0 Instruct. For local testing and memory-constrained devices — Bielik 4.5B. For edge and IoT — Bielik 1.5B. For content moderation — Bielik Guard 0.5B.

Bielik 11B v3.0 and 4.5B v3.0 support a context window of 8,192 tokens. Bielik 1.5B v3.0 supports 4,096 tokens. The BPE tokenizer is optimized for Polish and encodes Polish text with fewer tokens than general-purpose tokenizers.

Base is a raw language model — it predicts the next token without special instructions. Instruct is fine-tuned (SFT + RLHF) for dialogue, following instructions, and assistant tasks. For production deployments, always use the Instruct version.

Other

Bielik is a family of open language models optimized for the Polish language — from compact 1.5B models to the flagship 11B. The project is developed by Speakleash Foundation and released under the Apache 2.0 license.

Bielik is developed by SpeakLeash Foundation together with strategic, scientific, and technology partners, as well as a broad contributor community. The project was founded in Poland and is led by Polish AI researchers.

11 · Contact

Get in touch

Write to the right address. We reply by e-mail.