A general-purpose model from the SpeakLeash Foundation, trained on the ACK Cyfronet AGH supercomputers (Helios, Athena). A family from 1.5 bn to 11 bn parameters; version 11B v3 since January 2026, Bielik‑PL with a Polish tokenizer since April 2026. Apache 2.0 licence — commercial use free of charge
Knowledge
Bielik and PLLuM — Polish AI models in companies and public bodies
Two Polish language models (Polish LLMs), two different goals. Bielik (the SpeakLeash Foundation and ACK Cyfronet AGH) is a general-purpose model under Apache 2.0; PLLuM (the HIVE AI consortium led by NASK) is a family of models for public administration and official language, partly under open licences. Both can run in an organisation’s server room — below: licences, sizes, hardware, deployments and how to connect them through PROXY:AI without your own GPU. As of 3.10.2026
2Polish models
Apache 2.0Bielik’s licence
4–70 bnPLLuM parameters
3.10.2026data as of
01 / 08Bielik PLLuM What they share What they do not do on their own
What Bielik and PLLuM are
Polish language models exist because global models handle Polish inflection, official vocabulary and legal realities less well. Both projects are open — the weights can be downloaded and run in your own infrastructure — but they differ in purpose, licence and who stands behind them
A family of models from the HIVE AI consortium (NASK — lead, Wrocław University of Science and Technology, IPI PAN, IS PAN, OPI, University of Łódź, COI, ACK Cyfronet AGH), funded by the Ministry of Digital Affairs. Sizes 4, 8, 12 and 70 bn parameters; in May 2026 11 variants were released (base, instruct, chat) under licences allowing commercial use; older “‑nc” variants non-commercial only
Training on Polish texts (PLLuM: incl. Monitor Polski, the Journal of Laws, parliamentary transcripts), the documentation the AI Act requires for general-purpose models, availability on Hugging Face and in Ollama, the option to run on-premises
A model is not a system: it has no access policies, data masking, query log or answers from the organisation’s documents with a citation. That comes from the layer above the model — the gateway and the workspace (sections 05 and 06)
02 / 08
Bielik or PLLuM — a comparison
Two Polish model familiesSame criteria — different capabilities
Criterion
Bielik
PLLuM
Who builds it
SpeakLeash Foundation + ACK Cyfronet AGH; open-source community
HIVE AI consortium led by NASK; funded by the Ministry of Digital Affairs (approx. PLN 39.5 m in three tranches)
Purpose
General model: conversation, summarising, classification, writing — also outside administration
Official and legal language, letters, plain-language rewriting, RAG scenarios for administration
Sizes
1.5B · 4.5B · Minitron 7B · 11B (v3, Bielik‑PL)
4B · 8B · 12B · 70B (plus research models 8x7B‑nc, 24B from scratch, Polstral 24B)
Licence
Apache 2.0 — all versions, commercial use free of charge
Depends on the variant: Apache 2.0, Llama 3.3 licence or CC BY‑NC 4.0 (non-commercial only) — check per model
Where it runs
chat.bielik.ai, partner applications (incl. InPost), commercial deployments by integrators
mObywatel (assistant since 31.12.2025), Ministry of Digital Affairs, Gdynia (city search and BIP), Poznań (support for officials), PLLuM Chat
Hardware
From a laptop (1.5B) to one 24 GB card (11B); Minitron 7B on a 6–8 GB card
4B–12B on one card; 70B needs a server with several cards or an operator
For whom
Companies and public bodies that want one model for many tasks with no licence restrictions
Public bodies working on letters and regulations; companies after checking the variant’s licence
Licencje i rozmiary zmieniają się z każdą wersją — przed wdrożeniem sprawdzić kartę konkretnego wariantuŹródła w sekcji 07
| Criterion | Bielik | PLLuM |
|---|---|---|
| Who builds it | SpeakLeash Foundation + ACK Cyfronet AGH; open-source community | HIVE AI consortium led by NASK; funded by the Ministry of Digital Affairs (approx. PLN 39.5 m in three tranches) |
| Purpose | General model: conversation, summarising, classification, writing — also outside administration | Official and legal language, letters, plain-language rewriting, RAG scenarios for administration |
| Sizes | 1.5B · 4.5B · Minitron 7B · 11B (v3, Bielik‑PL) | 4B · 8B · 12B · 70B (plus research models 8x7B‑nc, 24B from scratch, Polstral 24B) |
| Licence | Apache 2.0 — all versions, commercial use free of charge | Depends on the variant: Apache 2.0, Llama 3.3 licence or CC BY‑NC 4.0 (non-commercial only) — check per model |
| Where it runs | chat.bielik.ai, partner applications (incl. InPost), commercial deployments by integrators | mObywatel (assistant since 31.12.2025), Ministry of Digital Affairs, Gdynia (city search and BIP), Poznań (support for officials), PLLuM Chat |
| Hardware | From a laptop (1.5B) to one 24 GB card (11B); Minitron 7B on a 6–8 GB card | 4B–12B on one card; 70B needs a server with several cards or an operator |
| For whom | Companies and public bodies that want one model for many tasks with no licence restrictions | Public bodies working on letters and regulations; companies after checking the variant’s licence |
03 / 08mObywatel — PLLuM City halls — PLLuM Applications and business — Bielik In SAIE — both
Where they already run
Both models have left the laboratory. Below: deployments publicly confirmed by their creators — no examples that cannot be checked
Since 31.12.2025 an assistant in the app answers questions about official matters and helps find applications and forms; it runs anonymously and keeps no chat history; over 10 m app users
Gdynia: integration with the city search engine and access to the BIP; Poznań: support for officials in finding information and answering questions; Ministry of Digital Affairs: internal assistant
chat.bielik.ai for users, integrations by the foundation’s partners (incl. InPost), deployments by integrators in law firms, clinics and HR departments — summarising and classifying documents
Through PROXY:AI both models are entries in the organisation’s model register: access policies, data masking and the query log work the same as for global models; TWIN:DESK uses them to answer from the organisation’s documents with a citation
10 m
That is how many users mObywatel has, where since 31.12.2025 a PLLuM-based assistant answers — anonymously, with no chat history. A Polish model in production, not in a laboratory
04 / 08
Hardware requirements
Three hardware tiers · indicative, 4–8 bit quantisationHardware sizing in the Initial Assessment
Laptop or an inexpensive card
6–8GB
Classification, summaries, simple letters. Lower quality, but the data never leaves the computer
Bielik 1.5BBielik 4.5BBielik Minitron 7BPLLuM 4B
Workstation or a server with one card
24GB
Full quality of 11–12B models for a team; Bielik 11B in 8 bits is about 12 GB, leaving room for context
Bielik 11B v3Bielik‑PL 11BPLLuM 8BPLLuM 12B
Server with several cards or an operator
≥ 2 × 80GB
PLLuM 70B in 4 bits is about 40 GB plus context — in practice a server, or a model at an operator in the EEA behind the same gateway
PLLuM 70BPLLuM 24B from scratchPolstral 24B
Rule: the model should fit in the card’s memory with room for context — otherwise speed drops, not qualityIndicative figures — verify on the model card
05 / 08
How to run a Polish model in an organisation
1Choose the variant
The task (letters, summaries, chat) and the hardware decide the size; check the licence of the specific PLLuM variant
2Enter it in the register
A model = an AI system in the organisation’s register: provider, licence, purpose, data — an AI Act duty regardless of the model being open
3Run it behind the gateway
On-premises (Ollama, vLLM) or at an operator in the EEA; access only through PROXY:AI — policies, masking, query log
4Connect the documents
RAG in TWIN:DESK: answers from the organisation’s documents with a source citation; Polish model, data in the organisation
5Measure and swap
Logs and costs per team; changing the model (Bielik ↔ PLLuM ↔ global) without changing the process
06 / 08
How it looks in SAIE
PROXY:AI
One gateway to more than 100 models — including Bielik and PLLuM run locally; access policies, data masking, budgets and a WORM log
TWIN:DESKAssistants on a Polish model with answers from the organisation’s documents and a source citation
Deployment and securityOn-premises or air-gap: the model and the data stay in the server room; hardware sizing in the Initial Assessment
07 / 08
Sources
- SpeakLeash Foundation — Bielik: models, Apache 2.0 licence, documentation
- Hugging Face — speakleash (Bielik 11B v3, Bielik‑PL, Minitron 7B)
- PLLuM — Polish Large Language Model (HIVE AI consortium, NASK)
- NASK, “From models to deployments — PLLuM on the path to real applications” (deployments: mObywatel, Gdynia, Poznań, MC)
- ITwiz, “11 new PLLuM models released” (21.05.2026)
- Regulation (EU) 2024/1689 (AI Act), Art. 3, 4, 53 (general-purpose models)
Informational material. Model sizes, versions and licences change — check the card of the specific variant before deployment. Hardware requirements are indicative. As of 3.10.2026
FAQAll questions →
Questions about Polish AI models
Bielik or PLLuM — which one for a public body?
For letters and official matters PLLuM is trained deliberately (Monitor Polski, the Journal of Laws, transcripts) and already runs in mObywatel and city halls. For general tasks — summaries, classification, chat — Bielik under Apache 2.0 has no licence restrictions. Behind a gateway you can have both and choose per task
Can a Polish model be used commercially?
Bielik — yes, all versions under Apache 2.0. PLLuM — it depends on the variant: the new models from May 2026 are under licences allowing commercial use, the older “‑nc” variants are non-commercial only. The licence is checked on the model card and entered in the system register
Do you need your own GPU server to run it?
Not always. Models up to 12 bn parameters run on one 24 GB card, Minitron 7B on a 6–8 GB card. Larger ones (PLLuM 70B) need a server with several cards — or a model at an operator in the EEA, available through the same gateway as local models
Does an open model exempt you from AI Act duties?
No. The deployer’s duties (register, literacy, prohibited practices, transparency) concern the use of the system, not its licence. An open model makes documentation easier — there is a model card and training data — but the register entry and the use policy are on the organisation’s side
Is a Polish model worse than global ones?
In general multilingual tasks — usually yes, because it is smaller. In Polish official and legal language PLLuM and Bielik often answer more accurately than global models of the same class, and the data does not leave the organisation. In practice both approaches are combined behind one gateway
Next
What else to explore
ChatGPT Enterprise, Copilot or your ownA comparison of three paths to AI in 7 dimensions and a hallucination quiz
TWIN:DESKAssistants on the organisation’s documents with a source citation
Polish AI Systems ActWhat companies and public bodies must prove from 28.10.2026
Deployment and securityOn-premises, EEA or air-gap — hardware sizing and architecture