After setting off a new round of storm in the field of open source large models, Chinese artificial intelligence start-up Moonshot AI is once again at the forefront of the geopolitical and technological game. According to the latest reports from people familiar with the matter and foreign media, after the recent official release of the new open source flagship model Kimi K3 with 2.8 trillion parameters, Dark Side of the Moon is currently planning to seek to obtain larger NVIDIA Blackwell architecture GPU computing resources to prepare and train its next-generation Kimi K4 large model.
This move is generally considered to be a continued push for cutting-edge computing power under the severe pressure of the United States’ increased semiconductor export controls to China. Prior to this, Trump administration officials and some U.S. AI companies had publicly accused Dark Side of the Moon, claiming that it transshipped and secretly held high-performance servers including Nvidia GB300 through a third country, and questioned its Kimi K3 model for distilling and distilling fine-tuning against similar U.S. flagship models. In this regard, industry analysts pointed out that considering Kimi K3's huge parameter scale, context window of up to 1 million Tokens and its unique hybrid expert (MoE) architecture optimization, the model has extremely high independent innovation in technical architecture and cannot be achieved simply by distilling external models in a short period of time.
Currently, with Kimi K3 showing amazing inference efficiency and computing power cost-effectiveness, technology giants including Microsoft have been revealed to be evaluating introducing it into enterprise-level services to significantly reduce inference operation costs. In order to further leapfrog the next-generation Kimi K4 model in terms of scale and intelligence, Dark Side of the Moon continues to leverage the potential of domestic chips through cross-data center cluster technology, while its desire for top computing power has also reached a new height.
At the same time, the United States is currently conducting internal discussions on whether to completely ban China’s open source weighting model. However, this potential ban has also encountered joint resistance from many global technology giants such as Nvidia, Microsoft, Meta, Dell, and Perplexity. The latter jointly sent a letter calling on the U.S. government to maintain the prosperity and development of the global open source AI ecosystem. The subsequent computing power layout of Dark Side of the Moon and the training progress of the next generation of large models will undoubtedly continue to become the focus of the global artificial intelligence industry.