Licence
Some “open” models restrict commercial use. We read that before choosing.
Models
Weights you can actually run, not a name on a slide.
We match an open-source model to the task, the language and the machine. Families that often come up are Llama, Mistral and Qwen. Those are examples of the category, not our ranking and not a promise of a score.
Polish
Answer quality in Polish is checked on your questions.
Size
A smaller model on a reliable server can beat the largest one that does not fit.
What we weigh up
- Whether the weights may be used by a commercial or public organisation on the data you have.
- Whether the model makes sense in Polish on your procedures, not on a test from the internet.
- How much memory a variant needs and how many people will ask at once. That leads to the GPU server.
- Whether answers must stick to documents. Then the model is only half of the layout, next to the index.
What we do not do with a model
We do not fine-tune weights on company data just to claim a “custom model”, when a document index is enough. Fine-tuning is needed less often than it seems, and it always raises the question of what may enter a training set.
We do not publish score tables. A number from someone else’s leaderboard does not say how a model will read your internal instruction.
Weight updates
Models change. Maintenance includes the decision of when to swap a version and how to check that the new weights did not break answers on your examples. That is part of maintenance, not a silent background update.