How Grab Built a Vision LLM to Scan Images

Grab built a custom 1B-parameter Vision LLM to extract information from Southeast Asian documents for eKYC verification. Starting with Qwen2-VL 2B, they progressed from LoRA fine-tuning to full parameter training, then built a lightweight model from scratch combining Qwen2-VL's vision encoder with Qwen2.5's compact language decoder. The four-stage training process included projector alignment, vision enhancement, language-specific visual training on synthetic OCR data, and task-specific fine-tuning. The final model achieved comparable accuracy to the 2B version while delivering 48-56% faster latency, addressing challenges with non-Latin scripts and diverse document formats across the region.