Moonshot AImodel systemcard
Kimi K3
#4of 27 · general
Standing
dot = this model · shaded = the field's middle 60% · track = full fieldGeneral
77.4#4/27
Math
41.2#16/27
Coding
72.4#4/27
Health
65.8#7/26
Engineering
81.2#4/17
Cyber
76.9#4/14
Analyst
50.0#6/11
Legal
52.1#10/26
Instruct
57.6#10/27
Agents
67.9#5/27
Professional
87.4#3/27
Multimodal
53.9#7/19
Hallucination
44.4#9/27
Speed
20.0#21/22
Value
63.7#5/27
Open
61.3#1/11
Measured
≈ imputed · amber dot = lab-self-reported among the sourcesGPQA Diamond93.3%#6
HLE51.5%#4
MMLU-Pro88.0%#11
LiveBench79.2%#5
SWE-bench V93.4%#4
LiveCodeBench87.2%#9
SciCode58.7%#3
CritPt23.4%#6
Terminal-Bench83.0%#3
APEX Agents39.3%#5
GDPval59.1%#5
Arena1486#5
AA Index59.7#4
Vals Index74.7#3
Epoch Capabilities Index157.0#6
Operations
$3 / $15$ per 1M tokens, in / out
$6blended
43tokens / sec
$0.35$ per solved task
1486arena elo · 3,559 votes
Series
Moonshot's models on the board, by releaseKimi K22026-01-27 · 28Kimi K32026-07-16 · 77