Kimi K2 Thinking

Kimi K2 Thinking is an open-source model that operates as a "thinking agent," reasoning step-by-step while using tools to achieve state-of-the-art performance on various benchmarks. It is capable of executing up to 200-300 sequential tool calls without human intervention, allowing it to solve complex problems across a wide range of tasks. The model uses Quantization-Aware Training (QAT) to support INT4 inference, which provides a roughly 2x improvement in generation speed.

View model card in Model Garden

Model ID	kimi-k2-thinking-maas
Launch stage	GA
Supported inputs & outputs	Inputs: Text, Documents Outputs: Text
Capabilities	Supported Function calling Structured output Thinking Not supported Batch predictions
Usage types	Supported Dynamic shared quota Not supported Fixed quota Provisioned Throughput
Versions	Kimi K2 Thinking Launch stage: GA Release date: Nov 13, 2025
Supported regions
	Model availability	United States global
	ML processing	United States Multi-region
Limits	global: Max output: 262144 Context length: 262144
Pricing	See Pricing.