MoE routing strategies shape model quality, balance and serving costs

Mixture-of-Experts models use a router for each token, selecting a small number of expert networks instead of running all of them. The choice affects model quality, workload balance and the communication costs of multi-GPU serving. Mixtral 8x7B, for example, exposes 47B parameters but activates 13B for each token.

The most common method is softmax top-k routing: Mixtral selects 2 of 8 experts, while gpt-oss-120b selects 4 of 128 and Qwen3-235B-A22B selects 8 of 128. Other designs include Switch-style top-1 routing with capacity limits, shared experts alongside routed experts, and node-limited selection. DeepSeek-V3 uses one shared and 256 routed experts, activating eight.

Sigmoid scoring with per-expert bias balancing is described as the leading approach for open frontier models in 2026. The bias changes according to expert load without adding auxiliary-loss gradients. Expert choice gives each expert a fixed token bucket but can use future tokens in causal models; one reported test found an abnormal loss reduction of about 10%. DeepSeek-V4 also uses hash routing in its first three MoE layers, while dynamic top-k varies the number of active experts.