We published an article today about our multilingual evaluation of smaller LLMs using the belebele dataset. Belebele is an easy-to-use dataset with multiple choice question in 122 langauge variants, we focussed on 8 European languages. The gemma-3 models performed best, but open models like OLLMo-2 showed quite good performance, too! Read more in the full article:
https://substack.com/home/post/p-172471752
About me
I work since more than 20 years as a developer, product manager and AI lead with language technologies. Starting with speech recognition and machine translation I now focus on education in semantic technologies and LLMs.
Check out my AI trainings.
Contact me and book your training.