Moonshot AI unveils Kimi K3, nearing US frontier
2026-07-17
Moonshot AI unveils Kimi K3, nearing US frontier
2.8 trillion parameters on benchmarks
Chinese AI startup Moonshot AI said on the 17th that it had developed a new AI model with performance close to the US frontier. The model is strong in autonomous software development and coding, and could narrow what has been seen as a gap leaving China AI nearly a year behind the United States.
According to Moonshot AI, the new model 'Kimi K3' has 2.8 trillion parameters. That is 75% more than DeepSeek's latest model 'V4', previously regarded as the highest-performing in China, and it is seen as one of the world's largest open-source models that disclose technical information.
Claims edge in coding and agents
The company said Kimi K3 is also strong in agent functions that autonomously handle a range of tasks. In comparisons it published, it said the model outperformed OpenAI's 'GPT-5.6Sol' and Anthropic's 'Fable 5' in both coding and agent categories.
Both models were launched only about a month ago. The Financial Times reported that Kimi K3 could overturn the common view that China AI trails the United States by 8 to 12 months.
Moonshot AI acknowledged that Kimi K3 falls short of the latest models from the two US companies in overall performance, including the quality of outputs, but said it beats existing models such as the previous-generation 'GPT-5.5' and 'Opus 4.8'.
Kimi K3 is also competitive on price. Compared with the cutting-edge models from OpenAI and Anthropic, its usage fee per certain number of tokens is 40% to 70% lower.
China competition in focus at Shanghai forum
The World Artificial Intelligence Conference opened in Shanghai on the 17th. Chinese President Xi Jinping attended the opening ceremony and delivered a speech for the first time. With China pressing ahead in a state-backed competition with the United States in AI, the announcement could further spur improvements in Chinese AI performance.
Enjoyed this article? Share it with your network!