Moonshot AI launches Kimi K2 Thinking, an open-source reasoning model pitched as a true thinking agent designed for complex multi-step reasoning and autonomous task execution, debuting from Beijing.

The model aims to outperform rivals like OpenAI and Anthropic in the reasoning space, emphasizing cognitive depth and tool use, with claims of automatically selecting hundreds of tools to complete tasks.

Benchmark results position Kimi K2 Thinking at the forefront of open-source agents, showing 44.9% on Humanity’s Last Exam, 60.2% on BrowseComp, and 71.3% on SWE-Bench Verified coding tasks.

The release signals a broader shift toward agentic AI, where autonomous systems define goals and execute tasks with limited human supervision, mirroring industry moves like Goldman Sachs’ autonomous AI coder deployment.

Kimi K2 Thinking relies on a trillion-parameter Mixture-of-Experts design and can perform 200–300 sequential tool calls per task, operating with minimal human intervention.

Moonshot argues superior agentic capabilities offer a more defensible edge than pure price competition amid ongoing market pressure.

The launch occurs amid China’s rapid AI development and global race discourse, with mentions of Nvidia’s Jensen Huang urging U.S. AI leadership and Chinese models being viable, cheaper alternatives.

An illustrative test shows K2 Thinking tracing an actor and film role through 20+ searches across Wikipedia, IMDb, and interviews to assemble a coherent answer.

Kimi K2 Thinking features a 256K context length for large-scale tasks and uses native INT4 quantization via Quantization-Aware Training to speed inference.

Industry coverage notes that DeepSeek released an AI model leveraging visual clues to broaden context, though independent verification of Moonshot’s figures was not provided.

Moonshot demonstrates K2 Thinking solving intricate problems through multi-step planning and tool use, including a geometry challenge with substantial nested reasoning.

Kimi K2 Thinking builds on the July-era K2 model, aiming to outpace OpenAI’s ChatGPT in autonomous task execution and tool usage, capable of selecting hundreds of tools automatically.