# What is a mixture-of-experts model?

A mixture-of-experts (MoE) model contains many specialist sub-networks but activates only a few per token, so Qwen 3.5 carries 397 billion total parameters but activates about 17 billion per token. This gives frontier-scale capability at mid-size inference cost, which is why most leading 2026 models are MoE.

MoE is the architecture behind the price collapse in open-weight inference: bases like Qwen 3.5 and Kimi K2.5 deliver near-frontier quality while serving at small-model economics.

For buyers, the practical takeaway is to read both numbers: total parameters describe capability ceiling, active parameters describe serving cost.

## See also

- [Compare MoE models](https://allocate.network/compare/kimi-k2-5-vs-qwen-3-5)

## Related terms

- [Inference](https://allocate.network/glossary/inference.md)
- [Open-weight models](https://allocate.network/glossary/open-weights.md)
- [Quantization](https://allocate.network/glossary/quantization.md)
- [MoE routing](https://allocate.network/glossary/moe-routing.md)

---

[HTML page](https://allocate.network/glossary/mixture-of-experts) · [Machine-readable catalog](https://allocate.network/catalog.json)