Kimi K2 Thinking: A Leap in Open-Source AI Agents

Moonshot AI has launched Kimi K2 Thinking, an open-source thinking model designed to tackle complex problems through step-by-step reasoning and tool integration. This new model sets performance benchmarks across reasoning, agentic search, coding, and writing tasks.

According to the announcement, Kimi K2 Thinking can perform up to 300 sequential tool calls without human intervention, showcasing its capacity for long-horizon problem-solving. This advancement is a result of scaling both thinking tokens and tool-calling steps during testing.

State-of-the-Art Performance

Kimi K2 Thinking achieves significant gains on key industry benchmarks. It scored 44.9% on Humanity's Last Exam (HLE) with tools, a benchmark assessing expert-level reasoning across numerous subjects. It also reached 60.2% on BrowseComp, a test for web browsing and information retrieval, and 71.3% on SWE-Bench Verified for coding.

Advanced Agentic Reasoning

The model demonstrates impressive problem-solving abilities, notably on HLE, where it achieved a state-of-the-art score of 44.9%. This involved complex planning and execution using tools like search, Python, and web browsing. It successfully solved a PhD-level math problem using 23 interleaved reasoning and tool calls.

Coding and Development Prowess

In coding tasks, Kimi K2 Thinking shows substantial improvements, with scores of 61.1% on SWE-Multilingual and 71.3% on SWE-Bench Verified. It exhibits notable performance gains in front-end development, translating ideas into functional products through multi-step workflows.

Web Browsing and Search Capabilities

On the BrowseComp benchmark, K2 Thinking achieved 60.2%, significantly outperforming the human baseline of 29.2%. This highlights its proficiency in goal-directed, web-based reasoning in dynamic environments. The model cycles through thinking, searching, browsing, and coding to refine hypotheses and construct answers.

General Capabilities Enhancement

K2 Thinking offers improvements in creative and practical writing, exhibiting stronger command of style, tone, and instruction adherence. Its responses to personal and emotional queries are marked by increased empathy and balanced perspectives.

Efficiency Through Quantization

To improve inference speed and reduce memory usage, K2 Thinking employs Quantization-Aware Training (QAT) with INT4 weight-only quantization on its MoE components. This results in a roughly 2x generation speed improvement with native INT4 inference, maintaining state-of-the-art performance.

Kimi K2 Thinking is currently available via chat mode on kimi.com and through its API, with a full agentic mode planned for release soon.