Kimi K2.5 is Moonshot AI's native multimodal model released in January 2026, delivering state-of-the-art visual coding capability and a self-directed agent swarm paradigm. Built on Kimi K2 with continued pretraining over approximately 15 trillion mixed visual and text tokens, it generates code from visual specifications — turning UI designs and video workflows into working implementations. Its agent swarm technology can self-direct up to 100 parallel sub-agents, each independently using tools to search, generate, analyze, and organize information, reducing execution time by up to 4.5× for complex research and writing tasks.
API|VisionReasoning|Open ModelModified MIT
Knowledge Cutoff
2026-02-02
Input → Output Format
Context Memory
262KIN66KOUT
AI Performance Evaluation
Arena Overall Score
1449
±5As of 2026-04-23
Overall Rank
No.36
24,213 Votes
Arena by Ability
Hard Prompts
1471±6No.37
Expert Knowledge
1487±14No.23
Instruction Following
1439±7No.35
Conversation Memory
1451±9No.40
Creative
1416±10No.42
Coding
1508±8No.24
Math
1477±14No.10
Arena by Occupation
Creative Writing
1425±8No.40
Social Sciences
1471±10No.30
Media
1421±9No.36
Business
1435±9No.52
Healthcare
1467±15No.44
Legal
1444±14No.57
Software
1493±7No.27
Mathematics
1483±17No.9
Source:Arena Intelligence
Overall
AA Intelligence Index
47%↑8%
LiveBench
69%↑9%
Reasoning & Math
GPQA Diamond
88%↑7%
HLE
29%↑12%
LB Reasoning
76%↑16%
LB Math
85%↑11%
LB Data
61%↑12%
Coding
AA Coding Index
40%↑5%
LB Coding
78%↑4%
LB Agentic
48%↑5%
TAU2
96%↑23%
TerminalBench
35%↑4%
SciCode
49%↑8%
Language & Instructions
IFBench
70%↑13%
AA-LCR
65%↑4%
Hallucination (HHEM)
14%↑4%
Factual (HHEM)
86%↓4%
LB Language
78%↑6%
LB IF
57%↑11%
Output Speed
Standard Mode
33tok/s↓49
First Output 1.42s
Reasoning Mode
32tok/s↓56
First Output 93.86s
