Engineering decision · 2026-08-22

Kimi K3: Self-Host or API
The 2.8T ledger

K3's weights went open on 07-27 (2.8T, the largest ever). The answer to 'deploy it yourself or call the API' depends on your monthly volume and compliance requirements — run the numbers first.

#kimi k3#self-host#open weights#cost comparison

Key points

2.8T

K3 total parameters

Stable LatentMoE, 896 experts with 16 active per token. Weights open-sourced 07-27.

$3 / $15

Official API price

Per million tokens input/output, cache hit $0.30, one flat price for the 1M window.

Self-hosting entry tier

Full-precision 2.8T is unrealistic; even after FP8/quantization you need multiple 8×H-class GPU nodes for a usable deployment.

1M

Context window

Same spec for API and weights. When self-hosting, the long-context KV cache is exactly where the memory goes.

What self-hosting 2.8T means

K3 is the largest open-weights model ever: a 2.8T-total MoE. The weights are free (open-sourced 07-27), but 'free to download' is not 'cheap to run' — even an FP8 deployment takes a multi-node GPU cluster, and stacking the 1M-context KV cache on top pushes self-hosting's fixed cost comfortably into five figures a month. Its sweet spot is extremely large, steady call volume, or hard compliance requirements that data never leaves your intranet.

Background

Requests surged after K3 went open, and Moonshot briefly paused consumer subscription expansion — the API-side heat says most people voted with their feet for the API. Meanwhile the DeepSeek price hike (08-17) made some people dust off the old 'self-host to save money' ledger, but at 2.8T scale the break-even point arrives very late.

Timeline

2026-07-16

K3 launches at WAIC, API available the same day.

2026-07-27

Weights go open; downloads and deployment requests surge.

2026-08

Community deployment reports appear: multi-node FP8 is the mainstream viable plan.

Confirmed vs watch out

Confirmed

The weights are open, the official API price is $3/$15, and the context is 1M — all from official channels. K3 has real call volume on QCode.

Watch out

'Open source = free' is an illusion: a 2.8T inference cluster costs five figures a month. Claims of 'running K3 on a single card' refer to heavily quantized demos, not production use.

Self-host vs API

Self-host

Fits: extremely large and steady monthly volume, hard compliance that data stays on the intranet, an existing GPU cluster. High fixed cost, low marginal cost.

API (QCode etc.)

Fits: everyone else. Zero fixed cost, elastic scaling, always the latest weights. Around $3/$15 at official price times the service rate.

How to decide

Compute the break-even: monthly fixed self-hosting cost ÷ API unit price = the critical monthly volume. At a 2.8T cluster's fixed cost, that threshold is very high. Then add two hard rules: hard compliance that data stays on the intranet → self-host only; highly variable volume → API only.

On QCode

kimi-k3 is on sale on QCode (with real calls in the last 30 days), official price times the service rate, no minimums and no subscription. Measure your real monthly volume on the API first, then evaluate whether self-hosting clears break-even — most people find the API is the end state.

FAQ

Are Kimi K3's weights really open?

Yes — open-sourced on Hugging Face on 2026-07-27: a 2.8T Stable LatentMoE, the largest open-weights model ever.

What hardware does self-hosting K3 need?

A production-grade FP8 deployment needs multiple 8-GPU H-class nodes; a single card or box only runs heavily quantized demos with no production value.

Is self-hosting cheaper than the API?

Only when monthly volume is extremely large and steady. A 2.8T cluster's fixed cost is five figures a month, so the break-even point is very high.

When is self-hosting mandatory?

Hard compliance scenarios where data must stay on the intranet (finance, government/enterprise intranets). There, cost isn't the main variable.

Do the API and the weights have the same capability?

Officially it's the same weights. The API side has mandatory reasoning (low/high/max tiers); self-hosting behaves the same.

How is K3 billed on QCode?

Pay-as-you-go at official price times the service rate; catalog prices follow the live /models listing.

Sources

Moonshot's official launch and open-source announcements (07-16/07-27), the Hugging Face moonshotai org page, community deployment reports (2026-08), QCode /models (08-22).

Run volume on the API first, then talk about building

kimi-k3 is on sale pay-as-you-go on QCode — let a real bill answer 'should we self-host.'