How much more do we pay for Polish in foreign AI models?
Author: Andrzej Albera, Responsible for strategy and the company's direction
Language models do not read words; they read tokens, which are pieces of words. Foreign models learned mostly from English, so an English word is usually one token for them, while a Polish word is two or three. So we pay more for Polish and fit less text into the model. We calculated what this costs and what a Polish tokenizer changes.
How we calculated it
Measurement: the preamble of the Polish Constitution in Polish and in English, with tokenizers compared by the creators of Bielik.
| Step | Value |
|---|---|
| Tokens per word, English, foreign model (Llama 3.2) | 1.17 |
| Tokens per word, Polish, the same model | 2.81 |
| Difference | 2.4 times more |
| Tokens for the same text (preamble): Polish vs English, foreign model | 653 vs 371, i.e. 76% more |
| Tokens per word, Polish, Bielik v3 Polish tokenizer | 1.62 |
| Result | 42% fewer tokens for the same Polish text |
What does this mean in practice? For every 10,000 tokens, a foreign model fits about 3,600 Polish words, and Bielik's Polish tokenizer about 6,200, three quarters more files, resolutions and letters at once.
Poland and the world
| Tokenizer | Tokens per Polish word | Comment |
|---|---|---|
| Bielik v3 before the change (Mistral-based) | 3.22 | starting point |
| Llama 3.2 | 2.81 | popular general-purpose model |
| Qwen3 | 2.69 | popular general-purpose model |
| EuroLLM | 1.88 | European model |
| Bielik v3 with a Polish tokenizer | 1.62 | fewest tokens per Polish word in this comparison |
How to read these numbers
- A token means money and memory. In models billed per token, Polish text costs more for the same content. In models run on your own servers, you pay in server time and a smaller context window.
- A Polish tokenizer is not a curiosity, it is a saving. The same server handles more letters, and the model sees more documents at once.
- The numbers are per word. A Polish text has noticeably fewer words than its English translation (about a quarter fewer in the preamble), so for the same text the difference is about 76%.
What this means for institutions
- Choose the model for the task, not the other way round. Letters, decisions and answers to residents are in Polish. TWIN:DESK, our Digital AI Desk, runs on Polish models such as Bielik and PLLuM, in the office's own server room. Read more about the models on our page Bielik and PLLuM in companies and public offices.
- See how much you really use. PROXY:AI records who uses AI and for what. Such a log shows where a Polish model is enough and where another one pays off.
- Calculate before deployment. The CDF methodology starts with an initial assessment, in which we check what texts an office processes and which model will be cheapest for them.
Polish does not have to be more expensive. This is what the Polish AI stack is about: Polish models, a Polish tokenizer and SAIE, which turns them into a tool for public offices. We wrote about the first step in Polish LLMs in public offices: the last mile.