If your company operates in Poland, its customers, contracts and internal documents are mostly in Polish. Bielik and PLLuM are Polish-language LLMs with open weights, trained specifically for Polish and the Polish cultural context. They are small enough to run on a single GPU, often on a GPU-equipped VPS. This article covers who builds them, which versions are current, how they are licensed and when they make more sense than a large global model.
Status as of November 2026. We do not quote benchmark scores: they change with every release, and what really matters is how a model performs on your own documents. Instead we focus on what can be verified: licences, sizes, hardware requirements and typical use cases.
Who builds the Polish-language LLMs Bielik and PLLuM
Bielik is developed by the SpeakLeash community under the Bielik.AI brand. The models and their variants (instruct models, quantised builds and the Bielik Guard safety classifiers) are published on the SpeakLeash organisation page on Hugging Face. The project’s next big event is Bielik Summit, announced for 26 November 2026, so keep an eye out for new releases there.
PLLuM (Polish Large Language Model) is a model family commissioned by Poland’s Ministry of Digital Affairs. In 2024 it was built by a consortium of research institutions; since 2025, development has continued under HIVE AI, an alliance of research organisations focused on open language technology for Polish public administration. The models are published under the ministry’s Hugging Face account (CYFRAGOVPL).
Current versions, licences and hardware requirements
The current Bielik generation is v3.0, released between November 2025 and February 2026. It includes an 11B model in two instruct variants. Bielik-PL-11B-v3.0-Instruct differs from Bielik-11B-v3.0-Instruct in its tokeniser, which has been replaced with APT4, optimised for Polish. There is also the smaller Bielik-Minitron-7B-v3.0 (about 7.5B parameters), distilled and pruned from the 11B model. The current PLLuM release, tagged “2512”, was published in January–February 2026. It includes PLLuM-4B and PLLuM-12B (base, instruct and chat; the 12B is built on Mistral-Nemo) plus the Llama-based Llama-PLLuM-8B and Llama-PLLuM-70B.
Memory in the table is parameters × 1 byte (8-bit) plus about 15% overhead. KV cache for context comes on top and grows with the number of concurrent users. We explain the method in How much VRAM does an LLM need.
| Model | Parameters | Licence | 8-bit memory (+ ~15%) | Example hardware |
|---|---|---|---|---|
| Bielik-11B-v3.0-Instruct / Bielik-PL-11B-v3.0-Instruct | 11B | Apache 2.0 | ~12.6 GB | 24 GB GPU (A10, A30); at 4-bit (~6.3 GB) also RTX A2000 12 GB |
| Bielik-Minitron-7B-v3.0-Instruct | 7.5B | Apache 2.0 | ~8.6 GB | 12 GB GPU (RTX A2000) |
| PLLuM-4B | 4B | Apache 2.0 | ~4.6 GB | 12 GB GPU, even at 16-bit (~9.2 GB) |
| PLLuM-12B | 12B | Apache 2.0 | ~13.8 GB | 24 GB GPU (A10, A30) |
| Llama-PLLuM-8B | 8B | Llama 3.1 | ~9.2 GB | 12–24 GB GPU |
| Llama-PLLuM-70B | 70B | Llama 3.1 | ~80.5 GB (~40 GB at 4-bit) | 4-bit: A40 48 GB; 8-bit: 2× A40 (96 GB) or RTX PRO 6000 96 GB |
Licences: Apache 2.0 or Llama 3.1
Bielik, PLLuM-4B and PLLuM-12B are licensed under Apache 2.0: you can use them commercially, modify and fine-tune them without extra permission. The Llama-PLLuM variants inherit the Llama 3.1 licence from Meta’s base model. It is a custom licence with additional conditions, including attribution, an acceptable use policy and a separate licence requirement for services above 700 million monthly active users. For a typical company this is not a blocker, but have your legal team read it before deployment.
Strengths: native Polish and low running costs
Polish-language LLMs are trained on large Polish text collections, so they usually stick to Polish grammar, inflection and register better than global models of similar size. They also tend to handle local context well: names of public bodies, administrative procedures, legal terminology. A Polish-optimised tokeniser such as APT4 typically splits Polish text into fewer tokens, which means more content fits into the same context window and less compute goes into each answer.
The second argument is cost. An 11–12B model at 8-bit takes about 12.6–13.8 GB, so it fits on a 24 GB A10 or A30 with room left for context. Those are exactly the configurations in our ready-made GPU packages, which despite the name run any model, not just DeepSeek. Bielik-Minitron-7B, or Bielik 11B at 4-bit, fits even on the entry package with an RTX A2000 12 GB. For multi-user production or the 70B model, a more powerful GPU server is the better choice.
Limitations: when a Polish model is not enough
In general comparisons, large global models (the biggest versions of DeepSeek, Qwen, GLM or Mistral) are usually stronger at complex reasoning, coding, maths and long multi-step tasks. Much of that comes down to scale: an 11B model will not replace one with hundreds of billions of parameters. Polish models are the right tool when:
- you work only or mainly with Polish text,
- the task is well defined: summarisation, classification, answering from documents, editing,
- cost matters and a single GPU is the budget,
- data sovereignty and model provenance matter, for example in the public sector or regulated industries.
In practice a hybrid approach often works best. A Polish model handles most simple Polish queries cheaply and quickly. Harder tasks, such as analysing several documents at once, generating code or working in English and German as well, go to a larger open model in the same infrastructure. Both can sit behind an OpenAI-compatible API, and a simple routing rule in your application decides which one answers. For the larger options, see Best open-source LLMs to self-host in 2026.
How to run Bielik and PLLuM on your own server
Both families are published in the standard Hugging Face format, so the usual inference servers support them:
- Ollama or llama.cpp with GGUF files is the quickest way to test on one machine. SpeakLeash publishes a GGUF build of Bielik-Minitron-7B, among others.
- vLLM is the production choice for multiple users: continuous batching, efficient KV-cache management and an OpenAI-compatible API. An FP8 build of Bielik-Minitron-7B is also available.
- SGLang is a good fit when you need structured output (such as JSON) in RAG and agent workflows.
We compare these tools in detail in Ollama vs vLLM vs SGLang vs NVIDIA NIM. Whichever you choose, get three things right: the chat template from the model card, a context-length limit that matches your available memory, and monitoring that tells you when the request queue starts to grow. For production, consider a safety filter too; SpeakLeash provides the Bielik Guard classifiers.
Use cases for companies operating in Poland
Customer service in Polish
A chatbot or agent-assist tool that answers from your knowledge base, terms and conditions and ticket history. Polish models are good at correct, polite Polish, and the modest hardware needs mean one GPU server can carry the traffic.
Document summarisation and triage
Contracts, minutes, meeting notes, correspondence with Polish authorities. An 11–12B model can summarise, extract dates and parties, or classify incoming letters. Combined with document search (RAG with a vector database), it becomes an internal knowledge assistant, and your data never leaves the company.
Public sector and its suppliers
PLLuM was designed with Polish public administration in mind, and hosting in a Polish data centre makes data-location requirements easier to meet. Typical tasks: summarising letters, drafting replies from templates, searching internal regulations and simplifying official language. Administrative decisions still require human review; the model only prepares a draft.
FAQ: frequently asked questions
Is Bielik free for commercial use?
Yes. Bielik v3.0 models are released under Apache 2.0, which allows commercial use, modification and fine-tuning with no licence fees. You only pay for the hardware the model runs on.
What is the difference between Bielik and PLLuM?
Bielik comes from the SpeakLeash community; PLLuM is commissioned by Poland’s Ministry of Digital Affairs and, since 2025, developed by HIVE AI. Both are tuned for Polish. PLLuM covers a wider size range (4B–70B) and targets public administration. The Llama-PLLuM variants use the Llama 3.1 licence; the others are Apache 2.0.
What hardware do I need to run Bielik?
Bielik 11B at 8-bit needs about 12.6 GB of GPU memory, so a 24 GB card such as an A10 or A30 is enough. At 4-bit, like Bielik-Minitron-7B, it fits on a 12 GB card. Add memory for context if many users work at once.
Is a Polish LLM better than GPT models for Polish text?
For complex reasoning, large global models are usually stronger. Polish models win on cost, on running entirely on your own server and on good Polish in well-defined tasks. The best test is a few dozen real examples from your own business.
Can I run PLLuM on a VPS?
Yes, if the VPS has a dedicated GPU. PLLuM-4B and PLLuM-12B fit on the 12–24 GB cards in our packages; Llama-PLLuM-70B needs 48 GB at 4-bit or about 96 GB at 8-bit.
Want to try Bielik or PLLuM on your own data? We will run the model on a VPS with a dedicated GPU or a GPU server in our Tier III data centre in Poland, fully managed 24/7 or with root access. Contact us and we will size the setup to your users and workload.