Kimi has unveiled its latest breakthrough in artificial intelligence with the launch of Kimi K2 Thinking, a state-of-the-art open-source model architected as a thinking agent. This new model represents a significant leap in autonomous problem-solving, capable of reasoning step-by-step while seamlessly using tools to tackle some of the most complex challenges in reasoning, coding, and web navigation.

K2 Thinking is built on the principle of "test-time scaling," where Kimi has scaled both the number of thinking tokens and tool-calling steps. This allows the model to execute up to 200-300 sequential tool calls without human intervention, maintaining coherent reasoning across hundreds of steps to solve problems that were previously out of reach for AI systems.

Record-Shattering Performance Across Key Benchmarks

The capabilities of K2 Thinking are not just theoretical; they are demonstrated by its state-of-the-art performance on a suite of demanding benchmarks. The model establishes new records in critical areas of AI evaluation:

1. Reasoning: On Humanity's Last Exam (HLE)—a rigorous benchmark of expert-level knowledge across over 100 subjects—K2 Thinking achieved a top score of 44.9% by actively leveraging search, Python, and web-browsing tools.
2. Coding: The model shows exceptional proficiency in software development, achieving 71.3% on SWE-Bench Verified and 61.1% on SWE-Multilingual, showcasing its ability to generalize across programming languages and agent scaffolds.
3. Web Navigation & Search: On the challenging BrowseComp benchmark, which tests the ability to find hard-to-locate web information, K2 Thinking scored 60.2%, significantly outperforming the human baseline of 29.2%.

A Leap Forward in Agentic Capabilities

K2 Thinking excels in long-horizon, multi-step tasks that require planning, execution, and adaptation. In a compelling demonstration of its deep reasoning, the model successfully solved a PhD-level mathematics problem through 23 interleaved reasoning and tool calls. This capacity for structured, long-form problem-solving makes it a powerful tool for academic research, complex data analysis, and software engineering.

Beyond technical tasks, K2 Thinking also delivers enhanced general capabilities:

1. Superior Writing: The model generates more vivid, imaginative creative writing and produces rigorous, logically coherent long-form content for academic and professional contexts.
2. Empathic Interaction: It handles personal and emotional queries with greater nuance, offering thoughtful, balanced, and actionable responses.
3. Designed for Efficiency and Scale

Recognizing the computational demands of advanced AI, Kimi has engineered K2 Thinking for high efficiency. Through Quantization-Aware Training (QAT), the model supports native INT4 weight-only quantization. This innovation provides a roughly 2x generation speed improvement while maintaining its benchmark-topping performance, making it a practical solution for large-scale deployment.

Availability

Kimi K2 Thinking is now available in chat mode on kimi.com, with its full agentic mode slated for release soon. Developers and enterprises can also access its capabilities through the Kimi K2 Thinking API.

This launch marks a pivotal moment in the evolution of AI agents, moving beyond simple task execution to true, multi-step reasoning and problem-solving. Kimi K2 Thinking is poised to become an indispensable tool for developers, researchers, and businesses aiming to solve complex, real-world challenges.