Model

Kimi K2 Thinking

OPEN

Moonshot AI · Kimi K2 family · released Nov 6, 2025

Open-weight 1T MoE reasoning agent with native INT4 weights and 256K context, chaining 200-300 sequential tool calls.

ReasoningCodingVisionFunction callingTool useAgentic

2177.2

Elo · rank #48

Parameters

1000B

Active params

32B (MoE)

Context

256K tokens

Architecture

Mixture-of-Experts, 384 experts (8 active + 1 shared/token), native INT4 quantization-aware training

License

Modified MIT License

Languages

—

API price (in/out)

$0.6 / $2.5

Modalities

text

Benchmark results

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

AIMEMath99.1%#6

best: GPT-5.2 · 100.0%

BrowseCompAgents60.2%#27

best: Kimi K3 · 91.2%

GPQA DiamondReasoning84.5%#58

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning44.9%#12

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning68.0%#26

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding83.1%#34

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

MMLU-ProKnowledge84.6%#30

best: Claude Fable 5 · 91.5%

MMLU-ReduxKnowledge94.4%#5

best: Qwen3.7-Max · 95.0%

SWE-bench VerifiedCoding71.3%#55

best: Claude Fable 5 · 95.0%

τ²-Bench TelecomAgents93.0%#17

best: Claude Opus 4.6 · 99.3%

Terminal-BenchCoding47.1%#2

best: Claude Opus 4.5 (High) · 59.3%

Run it locally

VRAM @ Q4

600 GB

VRAM @ FP16

2000 GB

Fits on (Q4)

Multi-node cluster required

Does not fit on a single RTX 4090 (24GB); native INT4 packs ~500-600GB combined RAM/VRAM, community reports only ~5-10 tok/s with heavy CPU/SSD offload

Quantizations

GGUF · INT4 (native QAT)

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-Thinking GHmoonshotai/Kimi-K2 DOCSmoonshotai.github.io/Kimi-K2/thinking.html APIAvailable via Moonshot AI API

API price $0.6/$2.5 · each benchmark row carries its own source badge (see methodology)