Today, I finished migrating the translation service for #jumble to a self-hosted setup. I initially looked into LibreTranslate, as it was an available option, and discovered it's an open-source translation model. However, after testing it, the quality was significantly lower than Jumble's official translation service—to the point of being unusable.
Then, I remembered that the translation performance from Ollama, which I use regularly, is quite impressive. So, I built a simple adapter that mimics the LibreTranslate API but delegates the actual translation work to Gemma3:27B running via Ollama. It's working perfectly.