Mixture of Experts is a neural network architecture where the model contains many "expert" subnetworks but only routes each input through a small subset of them. The routing is learned during training. The result is a model with very large total parameter count (often trillions) but much smaller active parameter count per token (typically tens of billions), allowing inference cost to scale sub-linearly with total model size.
MoE has become a dominant architectural choice for frontier language models because it lets labs scale capability without proportionally scaling inference cost. Mixtral (Mistral AI), DeepSeek V3, and several of the frontier labs' internal models use MoE architectures. Whether specific commercial models (Claude Opus, GPT-5, Gemini) use MoE is frequently unconfirmed, but industry consensus is that most do.
The tradeoff is training complexity. MoE models are harder to train stably than dense models, require specialized infrastructure for expert routing, and can be harder to quantize for edge deployment. For most enterprise deployments the architectural choice is invisible — what matters is the resulting cost-performance curve, which MoE has generally improved.