A developer built a multilingual food database for a calorie tracking app using a local LLM on an AMD 9070XT. Starting with English data, the LLM translated entries into Arabic and Spanish, generating 6,000 items in four hours. This approach avoids the English-only limitations common in nutrition apps.

After testing several models, the developer found Gemma3 12B to be the most reliable, despite a slower speed of ~80 tokens per second. The 7B variant was faster but had more errors. This selection highlights the trade-offs between translation quality and processing speed for local LLM applications.

The current database is a starting point, with plans to expand to 12+ languages and regional foods. A significant challenge remains with search, as fuzzy matching across multiple languages is complex. The developer is seeking community input on data sources and multilingual search improvements.