LLMcompare

Search

Search for a command to run...

A · open · Kimi

Kimi K2.5

Moonshot

B · open · Kimi

Kimi K3

Moonshot

0:12

overall capability wins (1 ties across 13 scored). Kimi K3 leads overall.

  • Reasoning0:6
  • Coding0:2
  • Tool use0:2
  • Pricing3:0
  • Speed2:0
  • Specs0:2

Verdict

Kimi K3 leads coding (widest gap: +39.3 on TermBench); Kimi K3 offers ~4.1x larger context (1.0M vs 256K); Kimi K2.5 is ~5.6x cheaper on a blended token basis than Kimi K3.

  • Kimi K3 leads coding (widest gap: +39.3 on TermBench)
  • Kimi K3 offers ~4.1x larger context (1.0M vs 256K)
  • Kimi K2.5 is ~5.6x cheaper on a blended token basis than Kimi K3

Cost for 1M in + 250K out

Illustrative chat workload at primary-provider list prices.

Kimi K2.5
$1.23
Kimi K3
$6.75

Kimi K2.5 is about 5.5x cheaper on this workload.

Price per 1M tokens

Full comparison

Organization

Moonshot

Moonshot

Family

Kimi

Kimi

License

Modified MIT

Kimi K3 License

Open weights

Yes

Yes

Release

Jan 27, 2026

Jul 16, 2026

Knowledge cutoff

-

-

API / provider

Moonshot

Moonshot

Modalities

text, image → text

text, image, video → text

Specs

Context window

256K

1.0M

B +793K

Max output

64K

64K

tie

Parameters

1.0T (32B act.)

2.8T (104B act.)

B +2K

Pricing

Input $/1M

$0.60

$3

A +$2.40

Output $/1M

$2.50

$15

A +$12.50

Blended $/1M (3∶1)

$1.07

$6

A +$4.92

Speed

tok/s

70

60

A +10

TTFT (s)

0.45

0.6

A +0.15

Reasoning

MMLU-Pro

87.1

—

—

GPQA Diamond

87.9

93.5

B +5.6 pts

Humanity's Last Exam

30.7

46.9

B +16.2 pts

AIME 2025

96.1

—

—

MATH-500

—

—

—

Humanity's Last Exam (with tools)

—

—

—

AA-Omniscience Accuracy

35.2

47.6

B +12.4 pts

AA-LCR v1.1

78.0

88.7

B +10.7 pts

CritPt

3.1

23.4

B +20.3 pts

MMMU-Pro

75.4

80.5

B +5.1 pts

IFBench

70.2

—

—

Chartography

—

—

—

Chartography (With Tools)

—

—

—

Coding

SWE-bench Verified

76.8

—

—

SWE-bench Pro

50.7

—

—

SWE-bench Multilingual

67.3

—

—

LiveCodeBench

85.0

—

—

Terminal-Bench 2.1

45.7

85.0

B +39.3 pts

Aider Polyglot

—

—

—

Terminal-Bench 3

—

—

—

BigCodeBench

—

—

—

SciCode

49.0

59.5

B +10.5 pts

CursorBench

—

60.8

—

SWE-Rebench

49.0

—

—

NL2Repo-Bench

—

—

—

DeepSWE

—

69.0

—

WebDev Arena

—

1,674

—

Terminal-Bench 4.0

—

—

—

CursorBench 4.0

—

—

—

FrontierCode v1.1 (Main)

—

—

—

Arena

LMArena Elo

—

1,485

—

Kimi K2.5

Context
256K / 64K out
Parameters
1.0T (32B act.)
Price
$0.60 / $2.50
Speed
70 tok/s · 0.45s TTFT
Modalities
text, image → text
License
Modified MIT

Prior Kimi open MoE generation - strong long-context agentic coding before K3.

Kimi K3

Context
1.0M / 64K out
Parameters
2.8T (104B act.)
Price
$3 / $15
Speed
60 tok/s · 0.6s TTFT
Modalities
text, image, video → text
License
Kimi K3 License

Largest open-weight model (2.8T MoE / 104B active) — official card reports strong reasoning and agentic coding results.

Benchmark charts

Winner bars are emphasized. Per-benchmark deltas sit above each chart.

Reasoning & knowledge

GPQA: B +5.6HLE: B +16.2CritPt: B +20.3AA-LCR: B +10.7Omniscience Acc.: B +12.4MMMU-Pro: B +5.1

Coding

TermBench: B +39.3SciCode: B +10.5

Tool use & function calling

Toolathlon V.: B +43.5τ³-Banking: B +31.8

Pick a different pair · Back to catalog