Saltar al contenido principal
AITier

Kimi K2

KimiKimiOpen WeightMIT · Uso Comercial

Descripción

Kimi K2-Instruct-0905 is the latest, most capable version of Kimi K2, achieving state-of-the-art performance in frontier knowledge, math, and coding among non-thinking models. This Mixture-of-Experts model features 32 billion activated parameters and 1 trillion total parameters, meticulously optimized for agentic tasks. Key features include enhanced agentic coding intelligence, extended context length to 256K tokens, and a hybrid architecture trained with MuonClip optimizer on 15.5T tokens. The model achieves 65.8% on SWE-bench Verified (single attempt), 47.3% on SWE-bench Multilingual, and excels at tool use with 70.6% on Tau2-retail. It is a reflex-grade model without long thinking, designed to act and execute complex tasks seamlessly.

Fecha de lanzamiento
2025-07-11
Parámetros
1.0T
Longitud del contexto
131K
Modalidades
text

Radar de capacidades

33
general
56
coding
68
reasoning
55
science
60
agents
0
multimodal

Rankings

Dominio#PosiciónPuntuaciónFuente
Ranking de codificación315
48.0
AA
Ranking general276
46.0
AA
Ciencia326
44.0
AA

Puntuaciones de benchmarks (LLM Stats)

(LLM Stats (zeroeval))

Chat

IFEvalGoogle Research (2023)89.8%Aut.
Tau2 Airline56.5%Aut.
Multi-Challenge54.1%Aut.

Communication

Tau2 Retail70.6%Aut.
Tau2 Telecom65.8%Aut.

Factuality

SimpleQA31.0%Aut.

General

C-Eval92.5%Aut.
MMLU89.5%Aut.
MultiPL-E85.7%Aut.
TriviaQA85.1%Aut.
CSimpleQA78.4%Aut.
ACEBench76.5%Aut.
Aider-Polyglot60.0%Aut.
SWE-bench Verified (Agentless)51.8%Aut.

Language

MMLU-Redux92.7%Aut.
MMLU-redux-2.090.2%Aut.
MMLU-Pro81.1%Aut.

Math

MATH-50097.4%Aut.
GSM8k97.3%Aut.
CBNSL95.6%Aut.
LiveBench76.4%Aut.
CNMO 202474.3%Aut.
MATH70.2%Aut.
AIME 202469.6%Aut.
PolyMath-en65.1%Aut.
AIME 202549.5%Aut.
HMMT 202538.8%Aut.

Reasoning

HumanEvalOpenAI (2021)93.3%Aut.
AutoLogi89.5%Aut.
ZebraLogic89.0%Aut.
HumanEval-ER81.1%Aut.
EvalPlus0.80 / 100Aut.
MuSR76.4%Aut.
GPQANYU + Cohere + Anthropic (2023)75.1%Aut.
SWE-bench Verified (Multiple Attempts)71.6%Aut.
SWE-bench Verified (Agentic Coding)65.8%Aut.
SWE-Bench Verified65.8%Aut.
SuperGPQA57.2%Aut.
LiveCodeBench v653.7%Aut.
LiveCodeBench53.7%Aut.
SWE-bench Multilingual47.3%Aut.
Terminal-Bench30.0%Aut.
OJBench27.1%Aut.
Terminus25.0%Aut.
Humanity's Last Exam4.7%Aut.

Índices de evaluación AA

(Artificial Analysis)
Math 500(OpenAI (2024), subset of Hendrycks et al. MATH (2021))
97.1
Mmlu Pro(TIGER-Lab (Univ. of Waterloo, Toronto, CMU, 2024))
82.4
Gpqa(NYU + Cohere + Anthropic (2023))
76.6
Aime(MAA (Mathematical Association of America))
69.3
Tau2(Sierra + U Toronto + Vector Institute (2025))
61.1
Math Index(Artificial Analysis)
57.0
Aime 25(MAA (Mathematical Association of America))
57.0
Livecodebench(UC Berkeley + MIT + Cornell (2024))
55.6
Lcr(Artificial Analysis)
53.0
Ifbench(Google Research (2023))
41.5
Terminalbench Hard(Stanford × Laude Institute (2026))
15.9
Intelligence Index(Artificial Analysis)
12.7
Hle(Center for AI Safety + Scale AI (2025))
7.4

Puntuaciones por categoría LLM Stats

(LLM Stats (zeroeval))
Instruction Following
90
Language
90
Structured Output
90
Legal
80
Finance
80
Healthcare
80
Biology
80
Chat
70
Math
70
Physics
70
Frontend Development
70
Chemistry
70
Reasoning
60
General
60
Communication
60
Economics
60
Tool Calling
60
Code
50
Factuality
30
Agents
20
Vision
0

Precios

Precio de entrada$0.57 / 1M tokens
Precio de salida$2.3 / 1M tokens
Precio mixto (3:1)$1.002 / 1M tokens

Velocidad

Tokens/seg0.0
Retraso del primer token0.00s
Tiempo hasta la respuesta0.00s

Ranking de Precios por Proveedor

Ranking de Precios por Proveedor

7 proveedores

Más barato: OpenCode ZenMás caro: DevPass (LLM Gateway)
ProveedorEntradaSalida
1OpenCode ZenMás barato
$0.4
$2.5
2FastRouter
$0.55
$2.2
3KimiPRINCIPAL
$0.57
$2.3
4OpenRouter
$0.57
$2.3
5Kilo Gateway
$0.57
$2.3
6Vercel AI Gateway
$0.57
$2.3
7DevPass (LLM Gateway)
$0.57
$2.3

Comparar precios entre diferentes proveedores de API para este modelo.

Fuentes externas

LLM StatsArtificial Analysis