月之暗面发布Kimi K2 Thinking模型

Core Insights - The article discusses the release of the Kimi K2 Thinking model by the company 月之暗面, highlighting its advanced capabilities in autonomous reasoning and tool usage [1] Group 1: Model Capabilities - Kimi K2 Thinking can autonomously perform 300 rounds of tool calls and multi-round reasoning without user intervention [1] - The model achieved a score of 44.9% in the HLE test, surpassing leading closed-source models such as GPT-5, Claude 4.5, and Grok 4 [1] - In the tool invocation test conducted by Artificial Analysis, Kimi K2 Thinking scored 93%, marking it as the highest score recorded by third-party institutions [1] Group 2: Market Performance - Kimi K2 Thinking ranks third in the intelligence index with a score of 67, following GPT-5 Code and GPT-5 [1] - The training cost for Kimi K2 Thinking was $4.6 million [1] - Within two days of its launch on the Hugging Face platform, the model was downloaded over 50,000 times [1]