460万美元训练出顶级大模型? 月之暗面杨植麟亲自回应

Core Insights - The latest model Kimi K2 Thinking has surpassed leading models like GPT-5 and Claude 4.5 in key benchmark tests, attracting global attention [1][3][6] Performance Breakthroughs - Kimi K2 Thinking utilizes a "super-sparse MoE" architecture, achieving significant efficiency with a trillion-parameter model [6] - The model has implemented "native INT4 quantization," enhancing inference speed and reducing hardware requirements [6] - It has achieved top scores in the TAU-Bench test, indicating a qualitative leap in "agent" tool invocation capabilities [6] Market Impact - Kimi K2 Thinking's download exceeded 50,000 within 48 hours of release, making it the most popular open-source model on Hugging Face [6] - The training cost of Kimi K2 is rumored to be only $4.6 million, significantly lower than OpenAI's GPT-5, which raises questions about the model's cost-effectiveness and market positioning [8] Community Response - The co-founders of the company addressed various concerns in an AMA session on Reddit, clarifying that the reported training cost of $4.6 million is inaccurate [10] - They emphasized that the focus on text models is a strategic decision rather than a short-term tactic for ranking [10] - The community has shown strong interest in Kimi's capabilities, with developers praising its performance and expressing concerns about its application in production environments [12][13] Future Developments - The company plans to enhance the model's general capabilities and is considering increasing the context length in future versions [12][13] - There are discussions about developing smaller models and improving the API request-based billing structure to better align with enterprise cost structures [13]