# How does routing work inside a mixture-of-experts model?

Inside a mixture-of-experts model, a small learned router picks which expert sub-networks process each token, so only a fraction of the model computes at a time. Qwen 3.5 carries 397 billion total parameters but activates about 17 billion per token; the router makes that selection at every layer.

This is routing inside the model, distinct from model routing across models. The MoE router is trained with the network and invisible to you; model routing is your platform-level choice of which model serves which workload.

The consequence for serving is memory: every expert must stay resident even though few fire per token, which is why MoE models need large multi-GPU memory footprints while charging mid-size prices per token.

## Related terms

- [Model routing](https://allocate.network/glossary/model-routing.md)
- [Mixture of experts](https://allocate.network/glossary/mixture-of-experts.md)
- [KV cache](https://allocate.network/glossary/kv-cache.md)

---

[HTML page](https://allocate.network/glossary/moe-routing) · [Machine-readable catalog](https://allocate.network/catalog.json)