Consensus Across Models
Sup.Ai is doing something really cool that I bet you guys could implement. They give the user the ability to submit to multiple models, which you have, but instead of just returning the results, they pass all three results through a model to evaluate the responses and construct a consensus using the best aspects of each response. Then, they return the final consensus to the user. The idea is that while one model might hallucinate or write poorly, it's unlikely that they all would at the same time. I would love to, for example, be able to combine Deepseek, GLM and Kimi K2 or something like that. The biggest trick, here, would be how you prompt the final model to generate a useful consensus, but if you pulled it off, it would be incredible. Another advantage is that if a model is unavailable or unwilling to complete the task, another model will, so the user never experiences a dropped response.
We’ve added this as Fusion. Enable multi-model mode, choose the models you want to compare, then switch on Fusion and choose a Synthesizer model. It combines the model responses into a final answer. This can help compare and reconcile responses, but the combined answer can still make mistakes.