Moonshot AI

Kimi K2 Instruct

Zero-eval
#1MATH-500
#1GSM8k
#1CBNSL
+23 more

by Moonshot AI

About

Kimi K2 Instruct is a language model developed by Moonshot AI. It achieves strong performance with an average score of 66.7% across 38 benchmarks. It excels particularly in MATH-500 (97.4%), GSM8k (97.3%), CBNSL (95.6%). It supports a 144K token context window for handling large documents. The model is available through 1 API provider. Released in 2025, it represents Moonshot AI's latest advancement in AI technology.

Pricing Range
Input (per 1M)$0.57 -$0.57
Output (per 1M)$2.29 -$2.29
Providers1
Timeline
AnnouncedJan 1, 2025
ReleasedJan 1, 2025
Specifications
Training Tokens15.5T
License & Family
License
Modified MIT License
Base ModelKimi K2 Base
Benchmark Performance Overview
Performance metrics and category breakdown

Overall Performance

38 benchmarks
Average Score
66.7%
Best Score
97.4%
High Performers (80%+)
12

Performance Metrics

Max Context Window
144.0K
Avg Throughput
45.0 tok/s
Avg Latency
1ms

Top Categories

reasoning
89.0%
math
86.6%
code
79.5%
roleplay
76.4%
general
62.0%
Benchmark Performance
Top benchmark scores with normalized values (0-100%)
Ranking Across Benchmarks
Position relative to other models on each benchmark

MATH-500

Rank #1 of 22
#1Kimi K2 Instruct
97.4%
#2DeepSeek-R1
97.3%
#3Llama 3.1 Nemotron Ultra 253B v1
97.0%
#4Llama-3.3 Nemotron Super 49B v1
96.6%

GSM8k

Rank #1 of 46
#1Kimi K2 Instruct
97.3%
#2o1
97.1%
#3GPT-4.5
97.0%
#4Llama 3.1 405B Instruct
96.8%

CBNSL

Rank #1 of 1
#1Kimi K2 Instruct
95.6%

HumanEval

Rank #3 of 62
#1GPT-5
93.4%
#2Claude 3.5 Sonnet
93.7%
#3Kimi K2 Instruct
93.3%
#4Qwen2.5-Coder 32B Instruct
92.7%
#5o1-mini
92.4%
#6Claude 3.5 Sonnet
92.0%

MMLU-Redux

Rank #4 of 13
#1DeepSeek-R1
92.9%
#2Qwen3-235B-A22B-Instruct-2507
93.1%
#3DeepSeek-R1-0528
93.4%
#4Kimi K2 Instruct
92.7%
#5DeepSeek-V3
89.1%
#6Qwen3 235B A22B
87.4%
#7Qwen2.5 72B Instruct
86.8%
All Benchmark Results for Kimi K2 Instruct
Complete list of benchmark scores with detailed information
MATH-500
MATH-500 benchmark
math
text
0.97
97.4%
Self-reported
GSM8k
GSM8k benchmark
math
text
0.97
97.3%
Self-reported
CBNSL
CBNSL benchmark
general
text
0.96
95.6%
Self-reported
HumanEval
HumanEval benchmark
code
text
0.93
93.3%
Self-reported
MMLU-Redux
MMLU-Redux benchmark
general
text
0.93
92.7%
Self-reported
IFEval
IFEval benchmark
code
text
0.90
89.8%
Self-reported
MMLU
MMLU benchmark
general
text
0.90
89.5%
Self-reported
AutoLogi
AutoLogi benchmark
general
text
0.90
89.5%
Self-reported
ZebraLogic
ZebraLogic benchmark
reasoning
text
0.89
89.0%
Self-reported
MultiPL-E
MultiPL-E benchmark
general
text
85.70
85.7%
Self-reported
Showing 1 to 10 of 38 benchmarks