Imagine a world where AI models could seamlessly collaborate, combining their unique strengths to solve complex problems. This isn't science fiction—it's the exciting reality of Large Language Model (LLM) assembling. Researchers are discovering that by combining multiple off-the-shelf LLMs, we can achieve far better results than using even the single most powerful model. But how do we efficiently harness the collective power of these diverse LLMs? Enter RouterDC, a novel approach that acts like a smart traffic controller, directing incoming queries to the most suitable LLM. Unlike traditional methods that simply average responses or rely on expensive reward models, RouterDC learns the optimal routing strategy using a dual contrastive learning approach. This means it identifies the top-performing LLMs for a given query and learns which model has the strongest skillset based on the query's content, thereby improving selection accuracy and efficiency. The results? Significant performance gains across a diverse range of tasks, from general knowledge and mathematical reasoning to code generation and understanding complex language. RouterDC is not only more accurate but also incredibly efficient, up to six times faster than traditional voting methods because it runs inference on only one selected LLM. This new research paves the way for smarter, faster, and more effective AI systems—ones that learn to leverage the strengths of diverse models, ultimately leading us closer to realizing the true potential of artificial intelligence.