Model

Kimi K2.5

OPEN

Moonshot AI · Kimi K2 family · released Jan 27, 2026

Predecessor to Kimi K2.6 (whose own docs explicitly call it 'Successor to K2.5'); Moonshot's first native-multimodal flagship, built via continual pretraining atop Kimi-K2-Base on ~15T mixed vision+text tokens. Introduced native image/video understanding and the 'Agent Swarm' research-preview multi-agent orchestration (self-directed, up to 100 sub-agents, ~1,500 coordinated tool-call steps), trained with Parallel-Agent Reinforcement Learning (PARL). Vendor claims near-frontier coding/reasoning/vision performance versus GPT-5.2 and Claude Opus 4.5-class closed models; independent Artificial Analysis Intelligence Index score was 38/100 (v4.1 methodology, ranked #14/95 open+closed models at time of measurement), corroborating strong but not universally frontier-topping performance. As of July 2026 it has been superseded as Moonshot's flagship by K2.6 but remains live and served via Moonshot's own API (platform.moonshot.ai) and third-party hosts (OpenRouter, Fireworks, etc.).

ReasoningCodingVisionFunction callingTool useAgentic

2253.5

Elo · rank #39

Parameters

1000B

Active params

32B (MoE)

Context

256K tokens

Architecture

1T-parameter MoE, 32B active/token, 384 experts (8 routed + 1 shared), 61 layers (1 dense), MLA attention, 160K vocab, SwiGLU; native multimodal via 400M-param MoonViT vision encoder

License

Modified MIT License

Languages

—

API price (in/out)

$0.6 / $3

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath96.1%#10

best: GPT-5.2 · 100.0%

Arena EloHuman preference1449#9

best: Claude Fable 5 · 1505

BFCL v3Agents64.5%#22

best: Hunyuan-A13B · 78.3%

BrowseCompAgents74.9%#20

best: Kimi K3 · 91.2%

CharXivVision77.5%#16

best: Muse Spark 1.1 (xhigh) · 88.4%

CyberGymCoding41.3%#9

best: GPT-5.6 · 84.5%

GPQA DiamondReasoning87.6%#40

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning50.2%#5

best: Claude Sonnet 5 · 57.4%

LiveCodeBenchCoding85.0%#24

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MathVisionVision84.2%#8

best: Seed 2.1 Pro · 92.6%

MathVistaVision90.1%#3

best: Seed 2.1 Pro · 90.7%

MMLU-ProKnowledge87.1%#14

best: Claude Fable 5 · 91.5%

MMMU-ProVision78.5%#14

best: Claude Opus 4.7 · 85.5%

MMMUVision84.3%#9

best: Claude Fable 5 · 89.3%

OSWorld-VerifiedAgents63.3%#17

best: Claude Fable 5 · 85.0%

SuperCLUEKnowledge61.5%#5

best: Baichuan 4 · 80.6%

SWE-bench ProCoding50.7%#37

best: Claude Fable 5 · 80.0%

SWE-bench VerifiedCoding76.8%#34

best: Claude Fable 5 · 95.0%

Terminal-Bench 2.0Coding50.8%#47

best: GPT-5.6 · 88.8%

Video-MMEVision87.4%#6

best: Seed 2.1 Pro · 89.2%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Kimi K2 family

Elo progression across releases

Kimi K2Jul 20251526.6 Kimi K2 ThinkingNov 20252177.2 Kimi K2.5Jan 20262253.5 Kimi K2.6Apr 20262517.3 Kimi K2.7 CodeJun 20262232.7

Resources

HFmoonshotai/Kimi-K2.5 GHMoonshotAI/Kimi-K2.5 DOCShttps://www.kimi.com/blog/kimi-k2-5 APIAvailable via Moonshot AI API

API price $0.6/$3 · each benchmark row carries its own source badge (see methodology)