How Grab Built a Vision LLM to Scan Images

Grab built a specialized Vision LLM to extract information from Southeast Asian identity documents, overcoming traditional OCR limitations with diverse languages and formats. After evaluating open-source models, they fine-tuned Qwen2-VL and ultimately created a custom 1-billion parameter model that achieved comparable accuracy to larger models while delivering 48-56% faster processing speeds for production eKYC verification.