Blog · Published · Updated
Bielik or PLLuM in a Company: Hardware and GDPR
Short answer
Bielik 11B or PLLuM 12B in 4-bit quantisation takes about 7 GB of GPU memory, so it fits on Hetzner's cheapest GPU server with 24 GB. That server costs €214 a month net, and with 8 hours of upkeep about €493. At documents with 3,000 input tokens and 1,000 output tokens, the Claude Sonnet 5.5 API is cheaper up to about 34,500 documents a month.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- analysis with an engineer and a free prototype
- €0
- team in ChatGPT or Claude, up to 10 people
- from €1,150
- automation, agent or application for one process
- from €3,000
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 4 min read
Statistics
Polish AI models in companies: statistics 2026
34,500
documents a month, from which a 24 GB GPU server with upkeep at €493 comes out cheaper than the Claude Sonnet 5.5 API. Scenario.
17.8%
According to Statistics Poland (GUS, 2026), 17.8% of Polish companies use AI technologies.
8.4%
According to Eurostat (2025), 8.4% of Polish companies with at least 10 employees used AI.
95.9%
According to Poland's data protection authority UODO (2026), 95.9% of the organisations surveyed do not consider themselves ready to deploy AI in line with GDPR.
41–58.5%
According to Poland's data protection authority UODO (2026), 41 to 58.5% of organisations see no link between AI tools and processing personal data.
Figures checked: 22 October 2026
How much GPU memory a model takes
As much as the weights weigh at the chosen precision, plus headroom for context. In bf16 each parameter takes 2 bytes, in 8-bit one, in 4-bit about half a byte. Add 15 to 25% to the result for context memory and driver overhead.
| Model | bf16 | 8-bit | 4-bit | On 24 GB? |
|---|---|---|---|---|
| Bielik 4.5B | ~9 GB | ~4.5 GB | ~2.7 GB | yes, even in bf16 |
| PLLuM 8B, Llama-PLLuM 8B | ~16 GB | ~8 GB | ~4.7 GB | 4-bit with room to spare |
| Bielik 11B, PLLuM 12B | ~22–24 GB | ~11–12 GB | ~6.5–7 GB | 4-bit with room to spare |
| PLLuM 8x7B | ~94 GB | ~47 GB | ~26–28 GB | no, it needs 48 GB or more |
| Llama-PLLuM 70B | ~140 GB | ~70 GB | ~40–43 GB | no, it needs 96 GB |
A 24 GB card handles the whole middle of both families in 4-bit. Only the 70B model needs a server with 96 GB.
Where the 48 GB in the PLLuM table comes from
The PLLuM infrastructure page gives the requirements for production serving: about 34 GB for the 8B model, 48 GB for 12B, 70 GB for 24B, 134 GB for 8×7B and 168 GB for 70B, plus 24 to 48 GB for search and the reranker. These are numbers without quantisation, with headroom for many users at once. Both sets of numbers are true and describe two different deployments.
Quantisation costs quality. No measurement of that cost has been published for Bielik and PLLuM, so the only sound procedure is a few dozen tasks from your process run in bf16 and in 4-bit. That is a day or two of work.
What a GPU server in the EU costs
Prices on 22 October 2026, with an IPv4 address, without the one-off setup fee.
Źródło: Hetzner, GPU servers (2026) ↗
The GEX45 has an RTX PRO 4000 Blackwell card with 24 GB and is available in Helsinki. The GEX63 and GEX131 have an RTX PRO 6000 Blackwell with 96 GB. The one-off setup fee is €209, €999 and €1,499 respectively. The €889 price from the December 2025 press release about the GEX131 is out of date. Add disk space for models and the index, backups, monitoring and a second machine for a test environment to the count.
When your own server is cheaper than the API
- GEX45 server
- €214 + 8 h of upkeep (€279) = €493
- GEX63 server for the 70B model
- €999 + upkeep = €1,278
- One document on the Claude Sonnet 5.5 API
- $0.016 (€0.0143)
Break-even: about 34,500 documents on the GEX45, 89,500 on the GEX63
Price of the GEX45 server with 24 GB of GPU memory from Hetzner's page on 22 October 2026, without the one-off €209 fee. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. The API at Anthropic's rates, at the ECB rate of 1.1205 USD per euro on 22 October 2026. The upkeep time and the document size are the scenario's assumptions. The local model's quality must be comparable, which you check with a test.
Below the threshold the API is simpler and cheaper. On Claude Haiku 5.5 a document costs twenty times less than on Sonnet 5.5, so the threshold rises to hundreds of thousands of documents. A Polish model on your own server usually wins on the requirement that data does not leave the company. We cover this in more depth in a Polish model or an API.
What about GDPR
Your own server room
no transfer
PLLuM assumes a fully on-premise deployment, with the data in the institution's infrastructure.
Hosting in the EU
a data processing agreement
Hetzner runs data centres in Falkenstein and Helsinki. An Article 28 GDPR agreement is a condition for the start.
Always with the company
the record and access
Legal basis, impact assessment, access control, log retention and training.
With Bielik, the terms of use add to the obligations: version 1.1 of 31 July 2025 forbids using the model to process any personal data, so HR documents and customer data need another model or a clarification with the SpeakLeash foundation. A model in the company's server room does not change who has the right to see HR documents. Isolation levels from the cloud to air-gap are described in on-prem and air-gap.
When not to host the model yourself
- You have a few hundred documents a month
- The task needs hard reasoning or code
- Nobody in the company will maintain the server, the drivers and the model
- You want to check first whether the idea works
- Data may be processed by a vendor with a data processing agreement in the EU
Every new model version means fresh quantisation, a test set and a check of the instructions. We deliver this work as part of private LLMs: deployment, tests on your tasks, monitoring and upkeep. We price the build from €3,000 net and upkeep individually. The variants' licences are covered in the article on PLLuM in a company.
