Imp-V1-3B: How A Tiny Model Is Beating Giants In Multimodal Llm Space