Chinese AI Model Kimi K3 Halts New Subscriptions

Kimi K3 was launched late last week and was trained on 2.8 trillion parameters, a measure of a model's scale, making it the largest open-source system released to date. Open-source models are published for others to examine and build upon, an approach Chinese developers have used to compete with American laboratories at a fraction of the cost. Benchmark tests placed Kimi K3 at the top of the Arena evaluation platform for front-end coding, and the New York Post reported that it outperformed Anthropic's Opus 4.8 and OpenAI's ChatGPT 5.5 on most coding tasks, though it remains behind Anthropic's most advanced Fable model. Lian Jye Su, chief analyst at the research group Omdia, said new releases routinely strain computing infrastructure, but argued that the episode showed Moonshot "does not have sufficient compute chips to serve the current surge in demand." He added that the company had probably underestimated the model's popularity, and that K3 is unusually demanding to run.

The bottleneck arrives as United States-led export controls restrict China's access to the most advanced chips, a persistent constraint for the country's AI firms. Kimi K3's release nonetheless unsettled American technology stocks, amid concerns that cheaper Chinese alternatives could erode the pricing power of established rivals. Moonshot, founded in 2023 by the researcher Yang Zhilin, has raised more than $5.5 billion and was valued at $30 billion in June; Reuters reported that it has approached Goldman Sachs and other banks about a possible listing in Hong Kong. Other Chinese developers are moving quickly as well: Alibaba has previewed Qwen3.8 Max, which it says was trained on 2.4 trillion parameters, and the startup Zhipu released a new system last month.