跳转到主要内容

Kimi K3 (max)

KimiKimi开源权重Kimi K3 License · 商用许可

描述

Kimi K3 is Moonshot AI's flagship open Mixture-of-Experts model for long-horizon coding, knowledge work, and reasoning. It has 2.8 trillion total parameters, activates 16 of 896 experts, and combines Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. The model supports a 1M-token context window, native image and video understanding, tool calling, structured output, and always-on reasoning with max thinking effort at launch.

发布日期
2026-07-16
参数规模
2.8T
上下文长度
1.0M
支持模态
image, text, video

能力雷达图

57
general
74
coding
93
reasoning
72
science
70
agents
80
multimodal

排行榜排名

领域#排名分数来源
智能体能力模型榜14
61.0
LS
代码能力榜2
98.0
AA
通用能力榜6
95.0
AA
多模态榜7
70.0
LS
科学能力8
93.0
AA

基准测试分数 (LLM Stats)

(LLM Stats (zeroeval))

Agents

AA-Briefcase1548.00 / 3000
Program Bench77.8%自报
DECK-Bench73.5%自报
Toolathlon73.2%自报
Kimi Code Bench v272.9%自报
OfficeQA Pro63.3%自报
Job Bench52.9%自报
MLS-Bench Lite48.3%自报
PostTrainBench36.6%自报
SpreadsheetBench 234.8%自报
AutomationBench30.8%自报

Code

FrontierSWE81.2%自报
DeepSWE 1.169.0%
DeepSWE67.5%自报
SWE-Marathon42.0%自报

Math

MathVision97.8%自报

Reasoning

GPQANYU + Cohere + Anthropic (2023)93.5%自报
CharXiv-R91.3%自报
BrowseCompOpenAI (2025)91.2%自报
Terminal-Bench 2.188.3%自报
MCP Atlas84.2%自报
Humanity's Last Exam56.0%自报
APEX-Agents37.6%自报

Search

DeepSearchQA95.0%自报

Vision

OmniDocBench91.1%自报
BabyVision85.7%自报
MMMU-Pro (with tools)83.4%自报
MMMU-Pro81.6%自报
PerceptionBench58.5%自报
WorldVQA51.0%自报
ZEROBench0.41 / 100自报

AA 评测指数

(Artificial Analysis)
Gpqa(NYU + Cohere + Anthropic (2023))
93.5
Terminalbench V2 1
85.0
Lcr(Artificial Analysis)
82.7
Coding Index(Artificial Analysis)
76.2
Intelligence Index(Artificial Analysis)
59.7
Scicode(UIUC + Argonne National Lab (2024))
58.7
Hle(Center for AI Safety + Scale AI (2025))
46.9
Tau Banking
46.0

LLM Stats 分类评分

(LLM Stats (zeroeval))
Productivity
100
Agents
92
Reasoning
78
General
60
Physics
90
Search
90
Biology
90
Chemistry
90
Math
80
Multimodal
80
Code
70
Tool Calling
70
Vision
70
Systems
40

定价

输入价格$3 / 1M tokens
输出价格$15 / 1M tokens
混合价格(3:1)$6 / 1M tokens
缓存读取价格$0.3 / 1M tokens

速度

Tokens/秒36.4
首Token延迟3.56s
首回答延迟58.53s

供应商价格排行

供应商价格排行

25 个供应商

最便宜: Together最贵: Tinfoil
供应商输入输出
1Together最便宜
$0
$0.00002
2Novita
$0
$0.00002
3Fireworks
$0
$0.00002
4Moonshot AI
$0
$0.00002
5CrofAI
$2
$8
6Requesty
$2.25
$11.25
7Vancine
$2.4
$12
8DevPass (LLM Gateway)
$2.83
$14.13
9DigitalOcean
$2.85
$14.25
10Kimi主要
$3
$15
11OpenCode Go
$3
$15
12Vivgrid
$3
$15
13GitHub Copilot
$3
$15
14OpenCode Zen
$3
$15
15AIHubMix
$3
$15
16Moonshot AI (China)
$3
$15
17Cortecs
$3
$14.999
18Neuralwatt
$3
$15
19Neon
$3
$15
20EmpirioLabs AI
$3
$15
21CoralBricks
$3
$15
22Charm Hyper
$3.2664
$16.332
23Venice AI
$3.75
$18.75
24GreenPT
$3.762
$18.81
25Tinfoil
$4
$20

比较该模型在不同 API 供应商之间的定价。

外部链接