Model

Kimi K2.6

OPEN

Moonshot AI · Kimi K2 family · released Apr 20, 2026

Successor to K2.5; Moonshot's flagship open-weight agentic/coding model, extends context to 256K, adds native image/video input, and scales its 'Agent Swarm' orchestration to 300 sub-agents / ~4,000 coordinated steps. Vendor claims top-of-open-weights on SWE-bench Verified and near-parity with GPT-5.5/Claude Opus 4.6-class closed models on several coding/agentic suites; independent Artificial Analysis Intelligence Index score was 44 (v4.1 methodology), and LMArena text arena placed it around #7 overall (Elo ~1522) as the top open-weight entrant at the time — both corroborate strong but not universally frontier-topping performance versus the more selective vendor benchmark set.

ReasoningCodingVisionFunction callingTool useAgentic

2517.3

Elo · rank #15

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)

$0.95 / $4

Modalities

text · vision · video

Benchmark results

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

Agents' Last ExamAgents21.7%#17

best: GPT-5.6 · 52.7%

AIMEMath96.4%#8

best: GPT-5.2 · 100.0%

Arena EloHuman preference1462#6

best: Claude Fable 5 · 1505

BrowseCompAgents83.2%#13

best: Kimi K3 · 91.2%

CharXivVision80.4%#10

best: Muse Spark 1.1 (xhigh) · 88.4%

GDPval-AAAgents1190.4#27

best: Claude Fable 5 · 1932

GPQA DiamondReasoning90.5%#21

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning54.0%#3

best: Claude Sonnet 5 · 57.4%

LiveCodeBenchCoding89.6%#7

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

MathVisionVision87.4%#5

best: Seed 2.1 Pro · 92.6%

MMMU-ProVision79.4%#10

best: Claude Opus 4.7 · 85.5%

MMMUVision86.3%#5

best: Claude Fable 5 · 89.3%

OSWorld-VerifiedAgents73.1%#11

best: Claude Fable 5 · 85.0%

SWE-bench ProCoding58.6%#12

best: Claude Fable 5 · 80.0%

SWE-bench VerifiedCoding80.2%#17

best: Claude Fable 5 · 95.0%

Terminal-Bench 2.0Coding66.7%#22

best: GPT-5.6 · 88.8%

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-K2.6 DOCShttps://www.kimi.com/blog/kimi-k2-6 APIAvailable via Moonshot AI API

API price $0.95/$4 · each benchmark row carries its own source badge (see methodology)