In recent years, open-source multi-modal AI models such as Qwen-Omni, LLaMA, and Gemma have achieved remarkable progress in language and vision understanding.

However, these models are still underperforming in Thai, struggling with linguistic nuances, reasoning patterns, and cultural or contextual understanding. This limitation restricts their adoption in Thai-speaking environments and hinders innovation in local AI applications.

The Thai MANGO project aims to close this performance gap by developing a Thai-centric multi-modal foundation model through two key strategies:

• Curating and fine-tuning open-source models (Qwen, LLaMA, Gemma) using authentic Thai data collected from diverse sources — not only from the internet, but also created by in-house engineers, contributed by community volunteers, and derived from educational materials and real courses.
• Benchmarking and evaluating these models with Thai academic standards, including TCAS (Thai university admission exams), to assess their comprehension, reasoning, and multi-modal understanding.

By advancing open and localized model development, Thai MANGO will support AI innovation in:

• Education — adaptive learning, tutoring, and content generation
• Travel & Tourism — multilingual assistants and cultural translation
• Call Center & Service — Thai-language conversational agents
• Research & Academia — enabling Thai-specific foundation models for future studies

Ultimately, Thai MANGO aspires to strengthen Thailand’s position in global AI research, empowering the ecosystem to build high-quality, culturally aligned, and accessible Thai AI systems.