How much cheaper are Kimi K2.6 calls on Cloudflare Workers AI now?

Cloudflare says the FP8 KV cache work drops Kimi K2.6 serving cost roughly 30% at 64 concurrent requests, on top of a 41% throughput gain and a context window that doubles from 686,000 to 1.37 million tokens. Existing Workers AI callers see the change with no code edits.

Do developers need to change any code?

No. The FP8 KV cache and INT4 weight compression are backend changes on Cloudflare Workers AI, so every existing call to Kimi K2.6 or GLM 5.2 picks up the longer context, higher throughput, and lower cost automatically. The API surface stays the same.

How much accuracy is lost from FP8 and INT4?

Cloudflare frames both as effectively lossless for these two models. New cache integrity checks add under 1% throughput and latency overhead as part of the quantization pipeline, and Cloudflare shipped the change after internal validation against the BF16 and 8-bit checkpoints.

Which other Workers AI models get the same treatment?

Cloudflare's post targets Kimi K2.6 and GLM 5.2 specifically because both are large mixture-of-experts models where quantization pays off most. Cloudflare has previously shipped BF16 exponent compression via Unweight for smaller LLMs, but this post covers only the two frontier open models.

Cloudflare · 2026-08-03 · major

Cloudflare Workers AI — FP8 doubles Kimi K2.6 context, INT4 cuts GLM 5.2 40%

Cloudflare quantized two of the biggest models on Workers AI. FP8 KV cache doubles Kimi K2.6 context from 686K to 1.37M tokens; INT4 weights shrink GLM 5.2 from 705 GB to 421 GB and speed decode up to 55%.

Cloudflare Workers AI header for Kimi and GLM quantization work

FP8 KV cache and INT4 weights land on Workers AI, doubling context and shrinking checkpoints for Kimi K2.6 and GLM 5.2.

Key specs

Kimi context	686K → 1.37M tokens
Glm checkpoint	705 GB → 421 GB

Quick facts

Vendor	Cloudflare
Service	Workers AI
Change	FP8 KV cache + INT4 weights
Kimi K2.6 context	686K → 1.37M tokens
Kimi K2.6 throughput	+41% at 64 concurrent
GLM 5.2 checkpoint	705 GB → 421 GB
GLM 5.2 decode speed	+55% at low concurrency

What is it?

Cloudflare quantized how Workers AI serves two of the biggest open MoE models. Kimi K2.6's KV cache now runs at FP8 instead of BF16 — that alone doubles the context window and cuts serving cost. GLM 5.2's weights compress from 8-bit to 4-bit, shrinking each checkpoint from 705 GB to 421 GB per GPU.

How does it work?

The FP8 KV cache trade halves per-token memory during attention, so the same GPUs hold twice as many tokens — Kimi K2.6 jumps from 686,000 to 1.37 million. GLM 5.2's INT4 weight compression frees GPU memory for 1.18 million tokens of KV cache per card and speeds decode up to 55% at low concurrency. Added cache integrity checks cost under 1% throughput.

Why does it matter?

Anyone calling Kimi K2.6 or GLM 5.2 on Workers AI now gets longer context, faster responses, and roughly 30% lower cost — with no code change. Two of the biggest open models get cheaper to serve at scale on the same hardware.

Who is it for?

developers using Workers AI

Frequently asked questions

How much cheaper are Kimi K2.6 calls on Cloudflare Workers AI now?: Cloudflare says the FP8 KV cache work drops Kimi K2.6 serving cost roughly 30% at 64 concurrent requests, on top of a 41% throughput gain and a context window that doubles from 686,000 to 1.37 million tokens. Existing Workers AI callers see the change with no code edits.
Do developers need to change any code?: No. The FP8 KV cache and INT4 weight compression are backend changes on Cloudflare Workers AI, so every existing call to Kimi K2.6 or GLM 5.2 picks up the longer context, higher throughput, and lower cost automatically. The API surface stays the same.
How much accuracy is lost from FP8 and INT4?: Cloudflare frames both as effectively lossless for these two models. New cache integrity checks add under 1% throughput and latency overhead as part of the quantization pipeline, and Cloudflare shipped the change after internal validation against the BF16 and 8-bit checkpoints.
Which other Workers AI models get the same treatment?: Cloudflare's post targets Kimi K2.6 and GLM 5.2 specifically because both are large mixture-of-experts models where quantization pays off most. Cloudflare has previously shipped BF16 exponent compression via Unweight for smaller LLMs, but this post covers only the two frontier open models.

Try it

https://developers.cloudflare.com/workers-ai/models/