Home › Models › Moonshot AI › Kimi K3

Kimi K3

Kimi K3 is the only live model here that accepts text, image, and video in one request. It is a 1T model with a 1M context window, built for dropping a whole repository or a long document into a single call. Published scores on record are 93.5% GPQA Diamond and 87.0% ProofBench v1.1, so it is also the route for formal proofs and hard science questions. Terminal-agent speed still favors the DeepSeek lanes; use K3 when the job is context width or video, not raw edit throughput.

Whole-repo ingestionFormal verificationProofsLong videoTri-modal inputLong-Context Repository Ingestion

Input$3/1M

Output$15/1M

Cached read$0.3/1M

Live telemetry is not published for this route yet.

Parameters1T

Context1M

Intelligence index30

Published benchmarks

Terminal-Bench 2.1 88.3%

GPQA Diamond 93.5%

ProofBench v1.1 87%

LiveCodeBench 88%approx

MathVision 90%approx

Call it

Other models from Moonshot AI

Kimi-2.7 & earlier 1T