Is it possible a pool of these running smaller sized models could somehow be cheaply combined into a MOE approach (as supposedly GPT-4 is) so create something c

Is it possible a pool of these running smaller sized models could somehow be cheaply combined into a MOE approach (as supposedly GPT-4 is) so create something cheap but higher quality?