There doesn’t seem to be much new leaderboard data for German LLM, but within the available data, Llama 3.1 Instruct appears to be quite good. That model is generally well-designed. For 12B, Mistral Nemo might be a good option. Additionally, since Qwen 2’s scores aren’t bad, Qwen 2.5, which made significant progress compared to Qwen 2, and its successor Qwen 3, may also be promising.
For multilingual models, Gemma 2 and Gemma 3 are generally excellent.
Since the Hub now has a feature to search for models by size, it would be even better to find a version fine-tuned for German, regardless of which model you choose.