Long context windows are a problem. I gave Qwen 2.5 70b a ~115k context and it took ~20min for the answer to finish. The upside of MoE models vs 70b+ models is that they have much more world knowledge.