Skip to content

Kimi K2.5

Moonshot AIOpen weightsNon-reasoningVerifiedReleased Jan 27, 2026Input: text, image
Compare with Grok 4.3Source scores ↗OpenRouter ↗
Overall score
53.0
Rank 86 of 222 · 44.4–61.7
Context window
256K
256,000 tokens
Price per 1M tokens
$0.60 in · $3.00 out
Source: BenchLM
Blended price
$1.20
3:1 input:output tokens
Output speed
—
Not measured yet

Category scores

Scores per capability area with the model's rank among ranked models in that category.

  1. 1Multilingual#887.3
  2. 2Instruction following#4885.8
  3. 3Math83.3
  4. 4Reasoning71.3
  5. 5Multimodal#3370.4
  6. 6Knowledge#7862.0
  7. 7Coding#6947.8
  8. 8Agentic & tool use#6733.9
Higher is better. #N is the model's rank in that category, where it is eligible to be ranked.

Benchmark results (67)

As reported in system cards, papers and public evaluations. Position compares this model with every other model reporting the same benchmark.

Multimodal

Design Arena Website1249#43 of 91
MMMU-Pro78.5#18 of 48
CharXiv77.5#33 of 42
MMMU84.3#5 of 21
MathVision84.2#15 of 20
SimpleVQA71.2#9 of 17
RealWorldQA81.0#11 of 16
VideoMMMU86.6#4 of 12
OmniDocBench 1.588.8#7 of 11
Video-MME (with subtitle)87.4#7 of 11
MLVU (M-Avg)85.0#7 of 9
CountBench94.1#4 of 8
RefCOCO (avg)87.8#6 of 8
Video-MME (w/o subtitle)83.2#4 of 7
CC-OCR79.7#4 of 7
AI2D_TEST90.8#5 of 7
MMVU80.4#3 of 7
DynaMath84.4#5 of 6
We-Math84.7#3 of 5
MMLongBench-Doc58.5#4 of 5
Video-MME87.4#2 of 2

Agentic & tool use

Gert Labs45.88#29 of 52
BrowseComp60.6#41 of 50
MCP Atlas29.5#42 of 42
Terminal-Bench 2.050.8#28 of 39
Claw-Eval52.3#31 of 36
OSWorld-Verified63.3#27 of 35
JobBench8.7#33 of 34
τ³-bench results65.7#12 of 23
Toolathlon27.8#21 of 22
DeepSearchQA77.1#12 of 20
BFCL v468.3#7 of 19
ResearchClawBench14.0#16 of 19
WideResearch72.7#12 of 16
QwenClawBench54.3#5 of 10
DeepPlanning14.4#7 of 7
MCP-Tasks59.1#5 of 5

Coding

SWE-bench Verified76.8#35 of 98
SWE-bench Pro50.7#61 of 69
LiveCodeBench v685.0#15 of 46
NL2Repo32.0#30 of 31
SciCode48.7#7 of 23
React Native Evals77.2#8 of 15
SWE-Rebench58.5#8 of 13
SWE-bench Verified*70.8#4 of 5

Knowledge

GPQA87.6#37 of 117
HLE30.1#44 of 69
MMLU-Pro87.1#7 of 67
SuperGPQA69.2#10 of 20
MMLU-Redux94.5#4 of 15
C-Eval94.0#1 of 10
MMLU-Pro (Arcee)87.1#2 of 5

Math

AIME 202596.1#4 of 38
AIME2695.8#7 of 29
HMMT Feb 202687.1#11 of 24
HMMT Feb 202595.4#4 of 13
HMMT Nov 202591.1#7 of 12
MMAnswerBench81.8#7 of 10
AIME25 (Arcee)96.3#2 of 5

Multilingual

SWE Multilingual73.0#30 of 46
MMLU-ProX82.3#8 of 16
INCLUDE83.3#6 of 10
NOVA-6356.0#6 of 7
PolyMath43.1#6 of 6
MAXIFE72.8#5 of 5

Reasoning

LongBench v261.0#8 of 17

Instruction following

IFEval93.9#7 of 53

Pricing by source

USD per 1M tokens.

SourceInputOutputCached in
List price (BenchLM)$0.60$3.00—
OpenRouter moonshotai/kimi-k2.5$0.50$2.50$0.15

Last published rate. Moonshot's Kimi API platform listed K2.5 at $0.60 input / $3.00 output per million tokens and said the same model supports both thinking and non-thinking modes. Moonshot's chat and Batch pricing pages, read September 25, 2026, no longer list it.

Arena ratings

Human-preference ratings from LMArena. Best-rated variant shown.

Text(thinking)1445#62 · 74.8K votes
WebDev (coding)(Thinking)1437#82 · 20.1K votes
Vision(Thinking)1269#46 · 28.2K votes
Documents(thinking)1430#34 · 21.6K votes

Open weights

From the Hugging Face repository.

Parameters
1.03T
Licence
other
Downloads (30 days)
317.2K

Speed

Measured by us through OpenRouter (median of trials).

Not measured yet.

Related models

Same family

Similar score