Nordlys Hypernova

I mean where to start, I started going into this rabbit hole of LLM routing around feb of 2025 and oh boy has it been a journey. First with the classic LLM router, then realized how is a LLM or me supposed to prompt which model is best for this prompt at a large case. Then we moved to classifiers, and classifying prompts is hard, and even more so if you want to generalize your router. Then, decided to pivot into a guess some sort of prompt optimizer or orchestrator, which didn't make sense, at this point we were just lost and this was July/August 2025. Then we came upon this paper, that just saved our lives, it opened up a whole new perspective for me and team. What if we just do good old traditional ML clustering and evaluate each LLM on the cluster, that means your router will scale with the size of the data, and generalize pretty well which is exactly what happened. Now, here we are building what we coined Mixture of Models models to blend together the strengths of all labs into one big mega model that is all knowing.