---
title: "How MoE Inference Differs From Dense Models"
description: "Why a mixture-of-experts model that selects about 13 billion parameters per token still needs over 90 GB for its weights, and how that splits the budget."
canonical_url: "https://fanout.sh/inference-eng/curriculum/19-01"
md_url: "https://fanout.sh/inference-eng/curriculum/19-01.md"
access: "pro-preview"
---

# How MoE Inference Differs From Dense Models

Why a mixture-of-experts model that selects about 13 billion parameters per token still needs over 90 GB for its weights, and how that splits the budget.

- Access: Fanout Pro (public preview)

## Public preview

A model can select a submodel of roughly 13 billion parameters per token and still require more than 90 GB for its weights. Buying memory from the smaller number leaves most of the model nowhere to live. Mixture-of-experts inference separates two budgets that were close together for the dense course model: the weights stored and the weights selected for a…

## In this lesson

- The model pin
- Resident weights
- Selected work and memory traffic
- The idle rank's obligation
- Practice: audit a fit claim

## Continue learning

Pro unlocks the complete lesson, video, practice, and code where available.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md