7 MoE Routing Strategies: From Top-K to Hashing, Why Mixtral and DeepSeek-V4 Took Different Design Paths

Sparse activation is the core of MoE (Mixture of Experts) models: every token is processed by only a small subset of experts. The routing strategy is exactly the set of rules that decides "who processes the token". Recently, Asif Razzaq from MarkTechPost published a guide explaining 7 MoE routing strategies, from the Top-2 softmax used by Mixtral to the hashing routing adopted by DeepSeek-V4. Below we walk through the key highlights.

Top-K Routing is the most common practice. For each token, the router calculates logits for all experts, applies softmax normalization, then selects the K experts with the highest scores. Mixtral uses Top-2 routing, while Switch Transformer uses Top-1. This approach is simple and effective, but it easily leads to uneven expert load. Many models add an auxiliary load balancing loss to ease this problem.

Expert-Choice Routing flips the perspective: it lets each expert pick the K tokens with the highest matching scores for itself. This ensures every expert processes the exact same number of tokens, achieving absolutely balanced workload. But tokens may be handled by different numbers of experts, and some tokens might end up not being selected by any expert at all.

Sigmoid Gating replaces softmax entirely. Softmax creates a competitive dynamic between experts, where the sum of all expert probabilities equals 1. Sigmoid makes independent judgment for each individual expert, allowing multiple experts to be activated at the same time, or even zero experts to be activated. DeepSeek-V2 uses Sigmoid gating, which is reported to improve training stability significantly.

Dynamic Routing does not fix the K value. The model dynamically decides how many experts to activate based on the difficulty of the token or the confidence level of the router: simple tokens use fewer experts, while complex tokens allocate more experts. This enables more fine-grained control over computation cost, but it is more complex to implement and much harder to train.

Hashing Routing takes a more radical approach: it directly maps tokens to experts via a hash function, completely abandoning the learned router mechanism. DeepSeek-V4 is reported to adopt this strategy. Its advantages are strong determinism, naturally balanced load (when the hash function is well designed), and no risk of routing collapse. The tradeoff is lost adaptivity: tokens can no longer select experts based on their own content.

The original article also introduces other variants such as Noisy Top-K and Expert Capacity Factor, which we will not expand on here. As we can see from all these methods, routing strategy design is essentially finding a balance between simplicity and load fairness. Top-K is simple but hard to balance load, Expert-Choice delivers perfect balance but may cause waste, and hashing routing is simple but sacrifices flexibility. Which strategy to choose depends on the model scale, training goals and hardware constraints.

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