Deploying a Polish LLM in Your Company: Bielik and PLLuM Hardware, Costs and GDPR (2026)

Running Bielik or PLLuM on your own infrastructure means a server with a GPU, an inference stack, a data processing agreement if you host with a provider, and a maintenance budget. Bielik 11B or PLLuM 12B at 4-bit fits in roughly 7 GB of VRAM, on a card that costs a few hundred euros a month. PLLuM's own table asks for about 48 GB for the same class of model, because it describes a different deployment.

Quick answer

Three numbers set the shape of the project: the model's footprint in VRAM, the monthly cost of the card, and the token volume you would otherwise buy from an API.

• VRAM. Rule of thumb: 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. Bielik 11B and PLLuM 12B: roughly 22–24 GB in bf16, roughly 6.5–7 GB at 4-bit.
• Hardware in the EU. Hetzner GEX131 (RTX PRO 6000 Blackwell Max-Q, 96 GB GDDR7, 256 GB DDR5 ECC) from €889.00 per month, its predecessor GEX130 at €838.00 per month plus a €79.00 one-time setup fee (Hetzner press releases, December 2025 and September 2024, excluding VAT).
• Licences. Bielik is Apache 2.0 across the family we checked. PLLuM has three regimes: the 4B and 12B variants under Apache 2.0, Llama-PLLuM 8B and 70B under the Llama 3.1 Community License, and every -nc- variant under CC-BY-NC-4.0, which rules out commercial use (HuggingFace model cards and pllum.org.pl, read 7 August 2026).
• The engineering work Syntalith delivers: AI automation from €3,500 net, AI apps and agents from €6,000 net, typical full builds €6,000–35,000 net, maintenance priced individually. The GPU server itself is a separate recurring line in the budget.

The start is free: a process scan (€0) is a 30-minute engineer call plus a written takeaway within two business days. If you want a portable document with architecture and a fixed quote before a hardware decision, the implementation specification is €1,200 net. If you commission the system build from us, we credit the specification fee toward the build. Which model to pick in the first place is the subject of our guide to Polish language models for business.

How much VRAM does a model actually take?

As much as the weights weigh in your chosen precision, plus headroom for context. This is arithmetic worth redoing yourself before anyone quotes you hardware.

Model weights are numbers. In bf16 each one takes 2 bytes, so an 11-billion-parameter model needs about 22 GB for weights alone. Quantization stores the same numbers at lower precision: 8-bit is one byte per parameter, 4-bit roughly half a byte. Add 15–25% on top for KV cache (the memory holding conversation context), activations and driver overhead.

The practical conclusion is simple: a 24 GB card, the cheapest dedicated GPU tier in Europe, handles the entire useful middle of both families at 4-bit. Only the 70B pushes a project onto hardware costing several hundred euros a month more.

Where does the gap between ~48 GB and ~7 GB come from?

From assumptions rather than an error. Both numbers are true and describe two different deployments.

The infrastructure section on pllum.org.pl (read 7 August 2026) states requirements for production serving: an 8B generator around 34 GB, a 12B generator around 48 GB, a 24B generator around 70 GB, the 8×7B around 134 GB and the 70B around 168 GB. Retrieval and reranking sit alongside that at 24–48 GB. The recommended cards are A10, V100, L40, A6000, A100, H100 and GH200.

Those figures assume unquantized serving with headroom for KV cache and concurrent requests. The page mentions neither quantization nor llama.cpp, Ollama or the GGUF format. That is a reasonable choice for a public institution buying hardware once and needing certainty about throughput with many simultaneous users. For a company that first wants to find out whether the model is any good on its own documents, the same table overstates the required budget by an order of magnitude.

The honest other side: quantization costs quality. For Bielik and PLLuM there is no published, dated measurement of that cost, so no vendor should promise you "only a few percent of loss". The only sensible procedure is your own evaluation set: a few dozen real tasks from your process, run against bf16 and against 4-bit, then compared. That is a day or two of work and settles the question for your case rather than someone else's benchmark.

What do you run it on?

A stack the model's authors actually support, because that turns weeks of integration into days.

• Ollama. First-class support for Bielik: SpeakLeash publishes to its own namespace at ollama.com/SpeakLeash. The shortest path from download to a working endpoint, good for a pilot and for an internal tool with moderate traffic.
• llama.cpp and GGUF. Official builds, including IQ variants with imatrix calibration. Bielik-11B-v3.0-Instruct ships 17 quantized variants targeting llama.cpp, LM Studio, Jan and Ollama (model card, read 7 August 2026). This is the CPU-plus-GPU route and the one for hardware outside the NVIDIA shelf.
• FP8 on Ada Lovelace and Hopper. Bielik-11B-v2.3-Instruct has official FP8 builds alongside GGUF (Q4_K_M, Q5_K_M, Q6_K, Q8_0) and 4-bit GPTQ. If you have an L40S or an H100, FP8 sits between bf16 quality and low memory use.
• Speculative decoding. Bielik-11B-v3.0-DFlash is a roughly 1B draft model published on 24 June 2026, meant purely to speed up the 11B. It does not work standalone.
• High-throughput servers. Bielik and PLLuM are standard architectures (Mistral, Mistral-Nemo, Llama 3.1, Qwen2.5 derivatives), so they are served by vLLM like their base models. None of the project pages names vLLM, though, so treat that as a property of the architecture rather than declared support.

What does a GPU server cost in the EU?

As much as the cheapest card that holds your model with room for context. Only prices the provider published officially belong in an article.

Prices exclude VAT. We deliberately publish neither an OVH hourly rate nor a GEX44 price here, because the available sources are either contradictory or second-hand. At this size of decision a figure taken second-hand can move the budget by tens of percent, so both belong in the provider's configurator on the day of purchase. It is also worth confirming whether the newer cards are available in the Warsaw region; hosting in France or Germany stays inside the EEA and is equivalent from a GDPR standpoint.

Add what the card's price tag leaves out: storage for models and the vector index, backup, monitoring, and a second machine if you need a test environment alongside production.

When does your own server beat an API?

Above a volume threshold you calculate yourself. Below it, you are paying to play sysadmin.

Modelled scenario. The numbers below illustrate the method rather than measuring any client. Substitute your own.

Monthly token volume =
  documents per month
  x (input tokens per document + output tokens per document)

Monthly API cost =
  token volume / 1,000,000
  x current provider rate per million tokens

Server break-even rate =
  monthly GPU server cost / (token volume / 1,000,000)

Assume 20,000 documents a month at 3,000 input tokens and 1,000 output tokens each. That is 80 million tokens a month. Against a server at €889 a month, the threshold is 889 / 80, or roughly €11.1 per million tokens. If your provider's blended rate is clearly below that, the API wins. If it is above, keep calculating.

We deliberately quote no API provider's prices here. Rates change several times a year, and an article from last quarter is a worse source on this than the pricing page open in your browser. Go to your provider's pricing page, take the rate for the model you actually use, and substitute it into the formula above.

Then add the labour to the server side: deployment, evaluations, monitoring and maintenance. The "server versus API" calculation is honest only once self-hosting carries a headcount or a maintenance contract too. We break that decision down further in a Polish LLM or an API model: how to choose.

What about GDPR?

On-premise solves the transfer question and leaves every other obligation in place.

Deployment on your own infrastructure. PLLuM states this directly as an architectural assumption: the deployment is entirely on-premise and data stays inside the institution's infrastructure (pllum.org.pl, read 7 August 2026). It is the strongest Polish-language reference for the "the data stays with us" argument, because it comes from a state project rather than a vendor. In that setup there is no processing on your behalf by anyone else, because there is no external party.

Hosting in the EU with an external provider. Here processing on your behalf does occur, and you need an Article 28 data processing agreement. Hetzner offers an optional DPA and EU data centres with no minimum contract term (product page, read 7 August 2026). That is a condition of entry rather than paperwork to tidy up after launch.

What your own server does not solve. Records of processing, a lawful basis, an impact assessment where one is required, access control, log retention and team training all stay with you. If the assistant reads HR files, a model running in your own server room does not change who is allowed to see them. We map the isolation levels from cloud to air-gap in on-prem and air-gap: levels of isolation, and the processing agreement itself in GDPR and DPA when deploying AI.

What the hardware quote does not show

Maintenance. This is where Bielik and PLLuM deployments most often come apart.

Release churn. PLLuM moved from the 2412 stamp through 2508 and 2512 to a batch of eleven models announced on 21 May 2026. Bielik: v2, v2.3, v3, then the variants with a dedicated Polish tokenizer on 15 April 2026, the DFlash draft model on 24 June 2026, and an announced v3.1. Every model swap means re-quantizing, re-running the evaluation set and checking that the prompts still behave. Without that, an upgrade is a gamble.

The generator is not the whole system. PLLuM's table budgets a separate 24–48 GB for retrieval and reranking. In practice that is a second component to maintain, with its own document indexing cycle and its own retrieval-quality regressions.

Self-hosting is not managed ML infrastructure. Configuration, drivers, CUDA updates and in-instance management sit with the customer, which cloud providers state plainly in their documentation. You buy hardware and network; running the model stays with you or with a partner who does it for you.

That is exactly the work Syntalith delivers: deploying the model, evaluations against your tasks, answer-quality monitoring and maintenance after launch. We price it like any other automation or app (from €3,500 and from €6,000 net), with maintenance priced individually, because it depends on how many models you keep and how often you swap them. Full rates are on the pricing page.

When NOT to self-host Bielik or PLLuM

Honestly: in most companies the first version of a process should not sit on your own GPU.

• Low volume. A few hundred documents a month is nowhere near the threshold in the calculation above. A fixed server cost loses to usage billing before you even count the engineering work.
• The task needs hard reasoning. Models at 4–12B parameters are not in the same weight class as frontier models on complex reasoning and code. An independent Oxido test published on 17 March 2026 (reported by Bankier.pl) covered 20 tasks across 10 categories; Google's model won, and Bielik and PLLuM landed in the lower part of the ranking. That is one test, but it is dated, independent and Polish, which makes it a better reference point than any vendor's claim.
• Nobody will maintain it. If there is no infrastructure person in the company and no maintenance partner in the budget, the model will age on the server alongside the drivers.
• You need one experiment. To find out whether an idea works at all, an API is faster and cheaper. Moving to Bielik or PLLuM in-house makes sense once the process is defined and the volume is known.
• The data is not sensitive. If nothing blocks processing with a provider that offers a DPA and EU hosting, the regulatory argument disappears and only the cost calculation is left.

If any of these fits your situation, we will say so at the scan, before you spend anything on hardware. What each of these models actually delivers we cover separately for PLLuM and for Bielik. For a lighter single-server setup, see also your own AI agent on a VPS.

How to start

The cheapest sensible first step is to calculate the volume and define the task. Buying a card comes later.

1. Book a free process scan: 30 minutes with an engineer plus a written takeaway within two business days. No sales deck, you talk to the engineer who would build it.
2. Prepare: how many documents a month, how long they are, what the model is supposed to do with them, what data sits inside, and whether there is a requirement that it never leaves your environment.
3. After the call you get a recommendation: a Polish model on a specific card and stack, a model over an API, an implementation specification, or an honest "at this volume your own server will not pay for itself".

Book a free process scan | See pricing | What an AI agent is

Related articles

• Polish language models PLLuM and Bielik for business
• A Polish LLM or an API model: how to choose
• PLLuM: deployment for business
• Bielik: a Polish open-source LLM for business
• On-prem and air-gap: levels of isolation