MoE Routing Strategies Explained: Top-K, Expert-Choice, Sigmoid Gating and Dynamic Routing (2026 Guide)

Why it matters

This article is a practical survey of Mixture-of-Experts routing, the mechanism that decides which expert networks process each token in modern sparse models. Because most teams inherit a router when they pick a checkpoint rather than choosing one, the piece walks through seven routing strategies found in production models and explains the tradeoffs each carries. It covers softmax top-k token choice, sparse single-expert routing with capacity buffers, shared-plus-routed expert designs, sigmoid gating with auxiliary-loss-free bias balancing, group- and node-limited routing for cheaper multi-node communication, expert choice routing, and hash routing with dynamic top-k. The guide highlights why routing matters beyond quality: routing decisions become network traffic under expert parallelism, shaping serving cost and load balance. It also flags a key caution that expert choice routing can leak information from future tokens in causal language models, which its makers measured as a suspiciously large drop in training loss. The recommended default for new pretraining runs is sigmoid scoring with bias balancing.