We present an end-to-end speech translation system for Mapudungun–Spanish developed for the IWSLT 2026 low-resource task. Building on the Canary-1B-v2 model, we apply parameter-efficient fine-tuning with a lightweight adapter and leverage an English-centered configuration as a proxy to enable translation. Experiments show that the system captures key phonetic patterns despite limited data, though it exhibits biases toward repetitive Spanish outputs. Our results highlight both the feasibility and the challenges of adapting multilingual foundation models to low-resource Indigenous languages.