PLLuM: Poland's State-Backed Language Model and What It Takes to Deploy It (2026)

PLLuM is a family of open Polish language models developed by a consortium led by NASK PIB and funded by Poland's Ministry of Digital Affairs. The models range from 4 to roughly 71 billion parameters and run on your own GPU server. Some variants ship under Apache 2.0, others are research-only, and that difference decides whether you may put the model into a company at all.

Quick answer

If you are weighing PLLuM as the engine behind an internal assistant or an automation, five facts settle most of it:

• Who runs it: a consortium led by NASK PIB, with Wroclaw University of Science and Technology, IPI PAN, the Institute of Slavic Studies PAN, OPI PIB, the University of Lodz and Cyfronet AGH. The current phase runs with HIVE AI alongside the Ministry of Digital Affairs. Models are published by the CYFRAGOVPL organisation on HuggingFace.
• Licence settles the question before quality does. The 4B and 12B variants in the 2512 series are Apache 2.0. Llama-PLLuM 8B and 70B carry the Llama 3.1 Community License. Anything with -nc- in the name is CC-BY-NC-4.0, meaning research and research only.
• Hardware: the project's own table runs from roughly 34 GB of GPU memory for an 8B generator to roughly 168 GB for the 70B, unquantized. At 4-bit the same models drop to somewhere between a few and a few dozen gigabytes.
• Benchmarks: PLLuM publishes none. There are no results tables on the model cards, verified 7 August 2026. Anyone claiming otherwise is quoting something that does not exist.
• Cost on the Syntalith side: AI automations from €3,500 net, AI apps and agents from €6,000 net, typical full builds €6,000–35,000 net, maintenance priced individually. GPU infrastructure is a separate line you calculate yourself.

The start is free: a process scan (€0) is a 30-minute engineer call plus a written takeaway within two business days. The wider market view of Poland's open models is in Polish language models PLLuM and Bielik for business.

Who builds PLLuM and who pays for it?

The consortium leader is NASK PIB, a state research institute. That is worth stating plainly, because a persistent error credits OPI PIB with leading the project; OPI is a member, NASK is the leader. The other members named on the model cards are Wroclaw University of Science and Technology, the Institute of Computer Science PAN, the Institute of Slavic Studies PAN, OPI PIB, the University of Lodz and Cyfronet AGH.

The current phase runs with HIVE AI under the Ministry of Digital Affairs. The pllum.org.pl site (read 7 August 2026) states a grant of PLN 5,999,102.67, funded by the Ministry of Digital Affairs under the programme "HIVE 2 - kontynuacja rozwoju i wdrażania dużego modelu językowego w administracji publicznej", grant number 4/WII/DBI/2026. The original project period given on the same page is March to December 2024. Total funding across all phases is not published, so no cumulative figure should be circulating in vendor decks.

Why does this matter to a buyer? Because a model maintained on public grant money carries a different risk profile from a commercial product. There is no SLA, no contractual support, no guarantee that the next version lands on a date. What you do get is a public licence and full access to the weights, so once downloaded the model keeps running on your side regardless of what happens to the project.

Which PLLuM variants may be used commercially?

This is the table to read before any architecture conversation. Picking the wrong variant is a legal error, and no amount of engineering fixes it afterwards.

Source: pllum.org.pl and the model cards in the CYFRAGOVPL organisation on HuggingFace, read 7 August 2026. The PLLuM-12B models are built on Mistral-Nemo-Base-2407. Naming is date-stamped in YYMM form: 2412 is December 2024, 2508 is August 2025, 2512 is December 2025. There is no such thing as "PLLuM 2.0"; the project calls the 2512 batch a second generation but versions by date.

In May 2026 the Polish outlet PurePC reported the release of eleven models at 4B, 8B, 12B and 70B in base, instruct and chat variants, alongside a statement that they were prepared for AI Act requirements. Treat that second point carefully: it is a project claim reported in the press, and no compliance certification stands behind it.

Why are the best-trained variants locked behind a licence?

Here sits the most interesting and least-reported tradeoff in the whole project. The model cards state that the openly licensed variants were continued-pretrained on roughly 28-30 billion tokens of Polish text, drawn from a copyright-clean subset. The -nc- variants saw roughly 150 billion tokens of Polish text, which is exactly why they carry a non-commercial licence (cards read 7 August 2026).

The consequence is simple and inconvenient. The variant with the deepest exposure to Polish is unavailable to a company. What you may actually deploy saw five times less Polish text. So when someone shows you a PLLuM demo with impressive Polish, the first question is which variant it is and whether you are allowed to run it.

For contrast, Bielik is Apache 2.0 across the whole family checked, with no non-commercial tier. We take that difference apart in the piece on Bielik as a Polish open-source LLM for business.

How much GPU memory does it really take?

The project publishes its own requirements table for an on-premise RAG deployment. We reproduce it faithfully, because it is the only official sizing guidance for this family (pllum.org.pl, Infrastructure section, read 7 August 2026):

The table assumes fp16/bf16 serving with headroom for KV cache and overhead. The page mentions neither quantization nor vLLM, Ollama or llama.cpp. For a company that is not serving hundreds of concurrent sessions, these numbers overstate what a real deployment needs.

Below is the same family costed with a rule of thumb you can redo yourself: bf16 is about 2 GB per billion parameters, 8-bit about 1 GB, 4-bit about 0.5-0.6 GB, plus 15-25% for KV cache, activations and CUDA context. This is arithmetic you can check yourself, and none of these rows is a measurement.

The gap between "48 GB per the ministry table" and "7 GB at 4-bit" for the same 12B model is the most practical number in this piece. The honest caveat attached to it: quantization costs quality, and for PLLuM nobody has published a measurement of how much. Anyone quoting you a specific percentage of quality lost at Q4 is guessing.

A modelled cost scenario, with your own numbers to substitute: a dedicated server with a 96 GB card at Hetzner (GEX131, RTX PRO 6000 Blackwell Max-Q) was listed from €889 per month net in the Hetzner press release of December 2025. That tier is enough for a 70B at 4-bit together with the retrieval stack. A 12B at 4-bit fits on the cheapest dedicated EU GPU class. Always compare that against your token spend at your actual volume, never per seat. The full hardware calculation is in deploying a Polish LLM: hardware, costs, GDPR.

How is PLLuM actually deployed?

The project itself describes three deployment patterns, and all three assume running on your own infrastructure with no data leaving it.

1. A citizen-facing chat on an institution's website. The model answers public questions grounded in that institution's content.
2. An internal employee assistant. Drafting correspondence, summarising documents, preparing responses. This is the pattern that transfers most easily into a company.
3. A knowledge assistant over internal documents (RAG). The model answers from specific passages of your documentation and points at the source.

That last pattern is what most companies ask about, and it explains the "retrieval + reranker" line in the hardware table. The generator is not the whole system: without solid retrieval, even a good model answers fluently and off-topic. How we assemble that in practice is in the piece on an AI knowledge assistant built on RAG.

The argument for keeping the model in-house comes here from the strongest possible source: the PLLuM page states outright that deployment is on-premise and data stays within the institution's infrastructure. It is hard to find a firmer footing for a GDPR conversation than a state project that designed it that way. Isolation levels, from EU hosting to a true air gap, are covered separately in on-prem and air gap: AI isolation levels.

Where is PLLuM running in public administration?

According to a NASK announcement, deployments include an internal assistant at the Ministry of Digital Affairs, the city offices of Gdynia and Poznan, and a virtual assistant in the mObywatel app built by the Central IT Centre (COI). NASK's stated 2026 roadmap covers expanding partnerships across public administration, adding more cities, advanced testing and evaluation, and work on Poland's first commercial text-and-image model.

Read that as a signal of project maturity. It falls short of being a production reference for a commercial buyer. Which of these are pilots and which are generally available has not been publicly confirmed, nor have user numbers. Claims about "millions of users" in sales decks have no basis in the available sources.

When is PLLuM the wrong choice?

Every honest piece about Polish open models needs this section, and few have it.

• The process needs hard reasoning or code. A 4-12 billion parameter model is not in the same weight class as the frontier models available through APIs. In an independent Oxido test reported by Bankier.pl in March 2026 (20 tasks across 10 categories: email writing, business advice, language correctness, Polish history and culture, legal and tax questions, marketing) Google's model won, and Bielik and PLLuM landed in the lower part of the ranking. The test's author, Marek Jeleśniański, called Bielik's placement a decent result given incomparably smaller resources. No public head-to-head of PLLuM against GPT, Claude or Gemini on reasoning or coding benchmarks exists, so the size of the gap cannot be stated, but nothing published contradicts its existence.
• You have no infrastructure team and do not want one. A GPU server, drivers, updates, monitoring, re-quantization on every new model version: this is continuous work. PLLuM's release cadence (2412, 2508, 2512, an eleven-model batch in May 2026) means repeated re-evaluation and prompt regression testing.
• Your volume is small. A GPU server costs the same at a hundred queries a month as at a hundred thousand. Below a certain threshold an API is simply cheaper. The calculation that settles it is in a Polish LLM or an API model: how to choose.
• Your constraint is not data location. If the requirement is GDPR compliance, a data processing agreement with an EU provider is often sufficient and cheaper. What exactly has to be signed is covered in GDPR and DPA when deploying AI.
• You need long context. PLLuM's model cards do not state a context window. Since it is undeclared, do not plan an architecture that assumes hundreds of thousands of tokens.

How Syntalith approaches PLLuM

We deploy PLLuM-family models on client infrastructure when the licence on the chosen variant and the constraint profile genuinely fit. We also work with API models, including Claude, and we say plainly that frontier models are stronger on hard reasoning. The choice is an engineering decision taken separately for each process. We have no public PLLuM deployment under a client name and will not pretend otherwise.

On a scan it looks straightforward: we take one process, count the volume, check what data moves through it, and tell you whether PLLuM on your own infrastructure holds up or an API comes out cheaper and better. No sales presentation, you talk to the engineer who would build it.

How to start

1. Book a free process scan: 30 minutes with an engineer plus a written takeaway within two business days.
2. Prepare: which process you want covered, how many queries a month, what data appears in it, and whether there is a requirement that it never leaves your environment.
3. After the call you get a recommendation: a specific PLLuM variant with its licence, an alternative such as Bielik or an API model, or an honest "this does not need a model of your own yet" if that is what the numbers say.

Book a free process scan | See pricing | Custom AI apps

Related articles

• Polish language models PLLuM and Bielik for business
• Bielik: a Polish open-source LLM for business
• Deploying a Polish LLM: hardware, costs, GDPR
• A Polish LLM or an API model: how to choose
• On-prem and air gap: AI isolation levels
• What an AI agent is: seven criteria