Singapore-based super app provider Grab has identified a significant gap: existing large language models (LLMs) fall short in comprehending the nuances of Southeast Asian languages, while OCR systems struggle with the wide variety of document templates prevalent in the region. To bridge this gap, Grab has set its sights on developing its own visual large language model by 2025. Leveraging Alibaba Cloud's Qwen2-VL2B model as a starting point, Grab has curated a dataset rich in Southeast Asian language content and constructed an internal synthetic data pipeline. Employing low-rank adaptation techniques for initial fine-tuning, followed by comprehensive full-parameter fine-tuning, the company has successfully crafted a lightweight yet powerful model. This innovative model surpasses the performance of various OCR tools and general-purpose LLMs. Looking ahead, Grab is committed to further advancing its in-house model development to broaden the horizons of its document processing capabilities.