← All questions
LLM FoundationsMedium· MoE· Routing· Inference

What is the difference between token-choice and expert-choice routing?

In token-choice routing, each token independently selects the K best experts, which leads to variable load across experts but is highly compatible with standard inference workflows. Conversely, expert-choice routing has each expert select a fixed number of tokens, which guarantees perfect expert utilization but results in variable coverage per token. Empirically, token-choice is preferred because expert-choice is generally incompatible with single-query serving requirements.

Code playground (Python)