Moonshot AI's Kimi K2 Thinking Model Released Publicly, Boasting 200-300 Step Agentic Reasoning

Moonshot AI has publicly released its advanced artificial intelligence model, Kimi K2 Thinking, a development announced by Frank Wang via social media. This new model is positioned as the latest iteration in open-source thinking models, designed to excel in complex, multi-step reasoning and dynamic tool invocation. The release marks a significant moment in the competitive AI landscape, particularly within the context of rapidly evolving large language models.

Kimi K2 Thinking builds upon the earlier K2 model and is noted for its ability to perform 200 to 300 sequential tool calls without human intervention, a capability that enhances its autonomous research, coding, and writing workflows. According to Moonshot AI, this feature allows the model to maintain coherent, goal-directed behavior over extended processes, surpassing many prior models in stable long-horizon agency. The model also incorporates native INT4 quantization, achieving a roughly 2x generation speed improvement while maintaining state-of-the-art performance.

The new model has demonstrated strong performance across various benchmarks, including Humanity's Last Exam (HLE) and BrowseComp, where it has set new state-of-the-art records. Moonshot AI, a Chinese company backed by Alibaba, developed Kimi K2 Thinking with a training cost of approximately $4.6 million. Its release under a Modified MIT License makes it accessible for broad use, with specific conditions for large commercial deployments.

Industry observers are noting Kimi K2 Thinking's potential to narrow the performance gap between open-source and proprietary AI models. The model’s advanced agentic capabilities and robust benchmark scores contribute to a growing trend of powerful AI innovations emerging from China, intensifying competition for established players like OpenAI and Anthropic. This release underscores the accelerating pace of AI development and the increasing sophistication of publicly available models.