Moonshot AImodel systemcard

Kimi K2

#17of 27 · general

k2.5/k2.6 point releases collapsed

1.0Tparams · MoE

15Ttraining tokens

MITlicense

2026-01-27released

weights ↗repo ↗

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

28.3#17/27

Math

60.9#9/27

Coding

54.4#11/27

Health

71.1#5/26

Cyber

53.8#7/14

Legal

68.0#3/26

Instruct

34.0#16/27

Agents

28.5#20/27

Professional

25.8#21/27

Multimodal

22.3#15/19

Hallucination

22.4#20/27

Value

41.0#17/27

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond90.5%#16

HLE37.7%#19

MMLU-Pro87.3%#14

AIME 202597.8%#2

LiveBench73.5%#15

SWE-bench V76.7%#21

LiveCodeBench70.2%#23

SciCode52.8%#13

CritPt8.0%#18

Terminal-Bench62.0%#20

APEX Agents23.4%#16

Arena1461#14

Vals Index55.2#17

Epoch Capabilities Index151.1#17

Operations

$0.17$ per solved task

1461arena elo · 37,350 votes

Series

Moonshot's models on the board, by release

Kimi K22026-01-27 · 28Kimi K32026-07-16 · 77

Neighbors

nearest capability fingerprints · deltas describe the neighbor

GLM-5.2→ stronger professional · stronger agent Mistral Medium→ weaker coding Grok 4.5→ stronger agent · broader GPT-5.6 Terra→ broader HY3→ weaker math · weaker coding