---
title: "KV Cache Quantization"
description: "What one cached token costs, how a quantized KV cache changes the reservation, the cache dtype flag and its values, and where the scales come from."
canonical_url: "https://fanout.sh/inference-eng/curriculum/11-05"
md_url: "https://fanout.sh/inference-eng/curriculum/11-05.md"
access: "pro-preview"
---

# KV Cache Quantization

What one cached token costs, how a quantized KV cache changes the reservation, the cache dtype flag and its values, and where the scales come from.

- Access: Fanout Pro (public preview)

## Public preview

Cached keys and values are charged by the token. They start at zero when a request arrives, grow by one position for every token it produces, and end with it, so what they cost a server is set by how much traffic is live and how far each context has run. Every price the quantized GEMM lesson settled behaves the…

## In this lesson

- What one cached token is made of
- The reservation recomputed
- The flag and its six values
- Where the scales come from
- What this release does not 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