Expert knowledge for digital decisions
Which open language models are suitable for the municipal corporation?
Short answer
The widespread families
| Family | Origin | Note |
|---|---|---|
| Llama | Meta | Widely used, many tools, own license terms |
| Qwen | Alibaba | Strong results, good multilingual support |
| Mistral | France | European, efficient for small sizes |
| Gemma | Compact, for simple tasks |
Read licenses carefully
"Open" does not automatically mean "freely usable." Some models are subject to licenses with restrictions for commercial use or above a certain user count. This must be checked before deployment – not after.
How to choose correctly
Online rankings say little about your use case. A practical approach:
- Collect twenty to fifty real cases from your daily work, each with the correct answer.
- Process them with two to three models.
- Compare: Accuracy rate, response time, costs.
This half-day effort is more valuable than any recommendation – including this one.
German as a criterion
Many models are predominantly trained on English texts. For German technical texts, official language, or Swiss spelling, significant differences emerge. Test with your own texts, not example sentences.
Important note on currentness
Models emerge and become outdated quickly. The mentioned families serve as orientation, not recommendations for a specific point in time. The decisive factor remains your own test with your texts – it can be completed in half a day and is more informative than any list.
Key facts
- Widely used
- Llama, Qwen, Mistral, Gemma
- More important than the name
- Test with own German texts
- Check before deployment
- License terms for commercial use
Sources
All external claims are backed by traceable sources.- 01
-
02
Artificial Intelligence Risk Management Framework: Generative AI Profile National Institute of Standards and Technology (NIST)