Poland already has its own LLMs. Now they need a road to the civil servant’s desk
Poland has something most of Europe is still working towards: its own open language models. Since December 2025, PLLuM has been answering citizens in mObywatel, the national government app, and in Polish-language tasks Bielik matches models several times its size. The state has paid for the engine and shown that it works. Now it is time for the last mile: the road from the model to the civil servant’s work computer.
This is the first post in the series AI in public administration: the last mile. In three parts we show how the public sector can make full use of Polish AI: today, models and money; in part 2, digital sovereignty and business continuity; in part 3, the civil servants who already work with AI.
Tuesday, 7:40 p.m., a municipal office
Anna is a clerk in a municipal office. She is fictional, but typical. She still has three replies to residents’ applications to write, and the deadline is tomorrow. She pastes a letter into a chatbot on her private phone and has a draft a minute later. She does it because she wants to finish on time and serve residents well, and the office has not yet given her a tool of its own.
Anna is not an exception. In a NASK survey, almost half of public sector employees said they use generative AI at work, the kind that writes text on request. That is good news: civil servants have discovered for themselves that AI saves time. Now it is worth moving that energy into a secure, official environment with a Polish model. How an office can take the lead is described in our Shadow AI guide.
PLLuM and Bielik: what Poland already has
- PLLuM was created in 2024 by a consortium led by Wrocław University of Science and Technology. Several variants are available under the open Apache 2.0 licence. The model is being developed further by the HIVE AI consortium led by NASK and funded by the Ministry of Digital Affairs. The state has invested several tens of millions of zlotys in it.
- Bielik was built bottom-up: the SpeakLeash community trained it together with the Cyfronet AGH Academic Computer Centre. Version 11B v3 is available under the open Apache 2.0 licence.
These are not lab curiosities. Since 31 December 2025, PLLuM has been working as the AI assistant in mObywatel and by 24 August 2026 had handled over 1.17 million messages in 730,000 conversations. The first offices also use it internally, and pilots are under way in Gdynia, Poznań, Łódź and Częstochowa, among others.
Read more about both models, their variants and uses in our guide Bielik and PLLuM in companies and public offices.
The last mile: from model to desk
A model is an engine. For a civil servant to use it every day, you need the whole car: access control, a usage log, masking of personal data, updates and a team to maintain it all. Today that gap is filled mainly by publicly available tools from foreign companies. Among Poles who use AI at work, almost 90% use publicly available chatbots and 55% do not tell their employer (KPMG, “Trust in AI 2025”).
Everyone involved has something to gain:
| Who | What they have today | What they gain on the last mile |
|---|---|---|
| Civil servant | A private chatbot account, often without training or a manager’s approval | An official tool with a Polish model, clear rules and training |
| Citizen | A Polish model in mObywatel and an unknown tool at the office counter | Certainty about where their data goes and who prepared the letter in their case |
| Office | No tool, no register, no view of how much AI is used | Control over data, an event log for the auditor and predictable costs |
A private account does not always mean the conversation trains someone else’s model. That depends on the product, the account type and the settings. But when a resident’s letter goes into a chatbot on a private account, the office does not know where the data is, how long it is kept or who can access it. An official environment with a Polish model gives that knowledge back to the office.
The state plans further investment in language models, in the order of PLN 20-30 million a year (PAP Samorząd). The biggest effect will come from shared, proven solutions that each of Poland’s nearly 2,500 municipalities can use, rather than thousands of separate projects.
How others did it
Countries that lead in public sector AI today started at the civil servant’s desk, not by building their own model:
- Singapore started in January 2023 with a hackathon. Today its government assistant Pair has over 98,000 monthly active users and is developed by a team of just a few people.
- The United Kingdom gave Copilot to 20,000 civil servants, who reported saving 26 minutes a day. A study at the Department for Work and Pensions, with a control group, found 19 minutes.
- Baden-Württemberg built the F13 assistant and opened its code. Neighbouring Saarland deployed it in four weeks.
Poland is in an exceptionally good position on this road. It already has what others built last: its own models and a working nationwide deployment. The Dutch spent EUR 13.5 million on their own model, GPT-NL, and are only now testing it. Poland can go straight to the step others started with: a secure tool for every civil servant.
When a Polish model is enough, and when to reach for a stronger one
In the hardest legal tasks, leading foreign models have the edge today. In a test based on questions from Polish legal professional exams, Gemini 3 Flash answers correctly in 100% of cases, Claude Opus 4.5 in 98%, GPT-5.2 in 86%, Bielik 11B in 73% and PLLuM 12B in 57% (PolishLawLLM Benchmark, an academic project without peer review).
Yet most everyday office work means summarising a letter, drafting a reply or finding a resolution. What counts here is good Polish, and in a Polish test suite Bielik 11B scores on a par with a Meta model more than six times its size (65.93 vs 65.49 points). The best model for an office is therefore chosen through trials on its own tasks, data classes and costs.
A sensible standard combines both worlds:
- sensitive data, such as PESEL numbers, residents’ cases or internal documents, goes to a Polish model running in the office;
- where data classification allows, a stronger model works through the same controlled gateway, for example PROXY:AI;
- which model works on which data is decided by the office’s policy, not by a clerk at 7:40 p.m.
The gap is closing fast. According to Epoch AI, the best open models are about four months behind the leading closed models.
The last mile can be travelled today
There is no need to invent it from scratch. We have travelled it with SAIE: TWIN:DESK is the AI digital desk that works with Polish models such as Bielik and PLLuM, PROXY:AI controls every query, and our CDF methodology guides the deployment from day one. SAIE is in daily use at a critical infrastructure company, and we have described that deployment step by step. The whole operates in line with ISO/IEC 42001. What SAIE offers municipalities, counties and central offices is described on our Public administration page.
A standard for the public sector is best described by requirements, not by a product name: Polish models, data kept in-house, a replaceable model, an event log, source code held by the contracting authority. Our deployments show that all these requirements can be met today.
Let’s talk at the 2nd Forum on AI in Public Administration
We will discuss how to travel the last mile in municipalities, counties and central offices at the 2nd Forum on AI in Public Administration in Wrocław (4-6 November 2026). We are a Special Partner of the Forum, and on 5 November we will give a talk titled “SAIE in a public office: an assistant answering from BIP, resolutions and procedures - with access control, usage logs and proof of compliance”. See Forum details.
AI training for public offices
A tool is half the success; the other half is people. We run AI workshops for civil servants. Our “AI in public administration” catalogue includes eight four-hour modules that can be ordered individually or as a full path. See our training offer.
Summary
Anna needs a tool on her work computer, powered by a Polish model. The citizen needs certainty about where their data goes. And the municipal secretary wants the money the state has already invested to work in their office too. Polish language models are ready. The last mile is the task for now.