(Yicai) Nov. 12 -- Moonshot AI’s chief executive officer said the USD4.6 million being reported as the cost to train Kimi K2 Thinking, the Chinese artificial intelligence startup’s new reasoning model, “is not an official number.

“It is hard to quantify the training cost because a major part is research and experiments,” Yang Zhilin also said during a question-and-answer session on US social media platform Reddit yesterday.

Training costs have become a benchmark for evaluating an AI model’s efficiency. USD4.6 million for Kimi K2 Thinking would be less than for DeepSeek V3 and OpenAI's GPT-3.

Moonshot has invested in the research, development, and updates of its open-source models over the past six months, releasing Kimi K2 Thinkin on Nov. 6, Yang noted. The Beijing-based firm will continue to adhere to its open source strategy, focusing on the application and optimization of the new model, he added.

Some users have criticized the long inference time of Kimi K2 Thinking and a gap between its leaderboard rankings and actual user experience. Yang said the model prioritizes absolute performance for now and will improve token efficiency and overall consistency in future versions.

More and more Chinese open-source AI models are being used overseas, with MiniMax M2, DeepSeek V3, GLM-4.6, DeepSeek V3.1, and DeepSeek-V3.2-Exp among the top 20 models on OpenRouter's AI rankings last week. Due to an interface issue on the platform, Kimi is only accessible via an application programming interface.

Continuous model updates and massive training workloads require substantial computing power, the Kimi team noted, adding that they use Nvidia's H800 graphics processing units equipped with InfiniBand, a computer networking standard used in high-performance computing and AI training, with each card used to its fullest potential.

While prioritizing text-based models, Moonshot will also advance the development of multimodal models, avoiding direct competition with market leaders such as OpenAI in the AI browser area. The company aims to set up a differentiated edge through architectural innovation, an open-source strategy, and cost control.

Editor: Martin Kadiev