Large language models like GPT-4 reach their limits when analyzing complex domains such as German tax law. With SteuerGPT, we present a domain-specific model that, despite its compact architecture (24B parameters), delivers significantly better results than much larger models like DeepSeek-R1 (70B) and GPT-4. Evaluated using a realistic benchmark based on actual tax law examinations, SteuerGPT demonstrates how targeted specialization using HPC infrastructure can lead to practical, real-world AI. Both the model and the benchmark are publicly available — a step toward trustworthy, tax law-informed AI. In this hands-on session, you have the opportunity to try SteuerGPT for yourself and get a feel for its capabilities and applications. The session is designed so that you can join at any time during the planned period.