Moonshot AI Reportedly Plans Larger Kimi K4 Model Amid US Chip Curbs

Chinese artificial intelligence firm Moonshot AI is in discussions to develop Kimi K4, a successor model that would be larger than its current version, according to a report from The Information. The plan highlights the company's ambition to push the boundaries of AI, even as it navigates stringent U.S. export controls on advanced semiconductor technology.

Navigating Chip Restrictions

The development of a more powerful model like Kimi K4 would necessitate access to a greater supply of advanced AI chips from companies like Nvidia, The Information reported, citing two people with knowledge of the matter. This presents a significant challenge, as current U.S. regulations are designed to prohibit such high-performance technology from reaching Chinese companies.

Moonshot is reportedly seeking to acquire more of Nvidia's latest Blackwell chips to train its future models. This underscores the intense global competition for the high-end computing resources essential for building large-scale AI.

Ad

Kimi K3's Training Origins

The company's current model, Kimi K3, is described as the world's largest open-source model with 2.8 trillion parameters. According to three people familiar with the matter cited in the report, it was trained using Nvidia chips, including the advanced Blackwell series.

Reports on the training location are varied. A White House senior official, Michael Kratsios, previously stated that Kimi K3 was trained at a data center in Thailand, a location that might fall outside the scope of U.S. restrictions. However, one source told The Information that the model was trained at least partly in China, noting that Chinese regulations on cross-border data transfer make it difficult to move the vast datasets required for pre-training to overseas facilities.