Kimi K3

kimi-k3

Open-weight frontier MoE model for reasoning, coding, and long-horizon agentic work

Intelligence: 43.6
Context: 1.04M
Training cutoff: --
Added: 2026-09-02
Inputs: Vision

Providers

3 routes serve Kimi K3, in the order requests are offered to them. Every figure is measured on its own route; where one has too little traffic in the selected window, it names what it is based on rather than borrowing the model's number.

Fireworks AI primary

accounts/fireworks/models/kimi-k3

Uptime

100.0%

Input / 1M: $3.00
Output / 1M: $15.00
TTFT: 4.79s fixed prompt · n=8
Output rate: 40.7 tok/s fixed prompt · n=8

1.04M context 1,048,576 max output cache $0.30 read / $3.00 write per 1M Vision ZDR Retention None Training No Provider terms, reviewed 2026-10-13warming · 30m

Prompt caching Automatic reuse Eligible Cache key Affinity hint Caller breakpoints None advertised

AkashML failover 1

moonshotai/Kimi-K3

Uptime

94.7%

Input / 1M: $1.20
Output / 1M: $14.00
TTFT: 1.05s fixed prompt · n=8
Output rate: 56.0 tok/s fixed prompt · n=8

1.04M context 1,048,576 max output cache $1.20 read / $1.20 write per 1M VisionRetention 30 days Training No Provider terms, reviewed 2026-10-15warming · 30m

Prompt caching Automatic reuse Eligible Cache key Affinity hint Caller breakpoints None advertised

Inco failover 2

kimi-k3:fast

Uptime

100.0%

Input / 1M: $6.00
Output / 1M: $30.00
TTFT: 0.76s fixed prompt · n=8
Output rate: 302.0 tok/s fixed prompt · n=8

1.04M context 1,048,576 max output cache $0.60 read / $6.00 write per 1M VisionRetention Yes, period not published Training No Provider terms, reviewed 2026-10-15warming · 30m

Prompt caching Automatic reuse Eligible Cache key Ignored Caller breakpoints None advertised

TTFT is the median time to the first visible answer token over real ModelRelay requests, with a fixed prompt standing in for models that have no traffic yet. Speed ranks by it. Intelligence is the Artificial Analysis Intelligence Index.