China's large models once again unleashed a "surprise." Moonshot AI officially open-sourced its strongest reasoning model to date - Kimi K2Thinking this Thursday. Not only did it outperform international closed-source giants such as GPT-5, Grok-4, and Claude 4.5 in the authoritative benchmark "Human Ultimate Exam" (HLE) with a score of 44.9%, but it also sparked a high-value AI revolution with a training cost of only 4.6 million US dollars (about 32.77 million Chinese yuan).
This achievement is particularly shocking due to its extreme cost control. According to CNBC citing sources, the investment in training Kimi K2Thinking is far less than 5.6 million US dollars for DeepSeek V3, and it stands in sharp contrast to OpenAI's research and development costs, which often reach billions of dollars. In the context of the AI industry generally falling into an "computing power arms race," Moonshot AI has proven through facts that stronger intelligence does not necessarily rely on higher investments.
More than performance, but also openness
Differently from most closed-source models, Moonshot AI chose to fully open-source: not only releasing model weights, but also publishing complete training scripts, data ratio plans, and evaluation toolchains, and explicitly allowing commercial use. This means developers can reproduce, fine-tune, or deploy enterprise-level applications at zero cost, greatly lowering the threshold for AI application. In the current context where the open-source ecosystem urgently needs high-quality Chinese reasoning models, Kimi K2Thinking undoubtedly fills a critical gap.
The term "reasoning model" emphasizes not just language generation, but also complex reasoning, multi-step planning, and knowledge integration capabilities. The Human Ultimate Exam (HLE) is a strict test measuring whether a model possesses human-like comprehensive intelligence, covering multidimensional challenges such as logic, common sense, mathematics, and ethics. Kimi K2's top ranking in this exam marks a new stage for domestic models, moving from "answering quickly" to "thinking deeply."
AIbase believes that Moonshot AI's move is not only a technical breakthrough, but also a redefinition of the global large model development path: the combination of open-source, efficiency, and commercial usability may drive AI from a "giant game" back to a "developer ecosystem." When a model exceeding GPT-5 can be trained with just 4.6 million dollars, the future of AI may be written by smarter and more practical Chinese teams.
