<h1>Kimi K2 Thinking: A New Milestone in Open-Source AI Models</h1>
Moonshot AI has launched Kimi K2 Thinking, an innovative open-source language model that is redefining the capabilities and economics of artificial intelligence. Developed in China, this one-trillion-parameter model utilizes a Mixture-of-Experts architecture and is natively designed for INT4 quantization, setting new standards in efficiency and cost-effectiveness.
Kimi K2 Thinking stands out by outperforming leading proprietary models like GPT-5 and Claude Sonnet 4.5 on key benchmarks. It boasts an impressive 44.9% score on Humanity’s Last Exam and excels in tasks requiring agentic capabilities by supporting 200-300 sequential tool calls. This unique feature allows it to tackle complex, multifaceted challenges with ease.
The model integrates advanced interleaved chain-of-thought reasoning and demonstrates exceptional competence in both reasoning and coding benchmarks, achieving a score of 71.3% on the SWE-Bench Verified test. This performance is enhanced by its quantization-aware training, which utilizes 4-bit deployment for a significant 2x improvement in inference speed.
Despite being a trillion-parameter giant, Kimi K2 is cost-effective, requiring less than $5 million to train. When operational, it costs $0.33 per million input tokens—a sharp contrast to GPT-5’s $1.25—making it financially accessible for broader deployment. However, potential users should note that significant commercial projects are subject to an attribution clause under the modified MIT license, especially when exceeding 100 million monthly active users or generating over $20 million in monthly revenue.
Kimi K2 is primarily hosted by Moonshot but does present challenges for self-hosting due to its substantial 594GB size. While it offers competitive pricing for its computational power—$0.60 per million input tokens and $2.50 per million output tokens—it remains a slower model operationally, processing at an eighteen tokens per second (TPS) standard.
This launch highlights the rapid advancements in AI emerging from Chinese labs, challenging Western companies and shifting the AI landscape both economically and geopolitically. Moonshot’s Kimi K2 Thinking underscores a pivotal move towards open-source models that provide robust, affordable solutions for complex AI applications.

Collections

Moonshot AI released Kimi K2 Thinking, a one-trillion-parameter open-source language model that outperforms GPT-5 and Claude Sonnet 4.5 on key benchmarks. The model uses Mixture-of-Experts architecture with INT4 quantization, achieving 71.3% on SWE-Bench Verified while supporting 200-300 sequential tool calls. It costs under $5 million to train and $0.33 per million input tokens to run—significantly cheaper than GPT-5's $1.25. The 594GB model operates at 18 tokens per second and requires attribution for large commercial deployments under its modified MIT license.

Exploring the Capabilities of Kimi K2 Thinking: Moonshot AI's Open Model Innovation