AI Hub
All models
LLMOpen weights

Kimi K2 Thinking

Moonshot AI

Updated May 21, 2026 · Benchmarks via Artificial Analysis, specs & pricing via OpenRouter · Methodology · Report an error

Superseded by Kimi K2.5

Specifications

Type
LLM
Access
Open weights
Released
November 6, 2025
License
Modified MIT
Parameters
1T (32B active)
Context window
262,144 tokens
Output speed
100 tok/s
Latency (TTFT)
1s
Input
Text
Output
Text
API pricing
$0.60 in · $2.50 out / 1M tokens
Capabilities
Function callingStructured output

Benchmarks

Reasoning

Math

General

Humanity's Last Exam (w/ tools)
44.9
MMLU-Pro
84.8
Humanity’s Last Exam
22.3

Other

Humanity's Last Exam (w/ tools)
44.9

Our take

Reasoning/agentic variant that thinks while using tools across hundreds of sequential calls, shipping in native INT4 via quantization-aware training.

Related research

Links

Compare Kimi K2 Thinking with

See all Kimi K2 Thinking alternatives →