(Yicai) July 22 -- The friendly rivalry between Moonshot AI and Elon Musk rumbles on, with the Chinese artificial intelligence startup pushing back after the US tech baron said SpaceXAI's Grok 4.5 is expected to exceed the performance of Moonshot AI’s latest model.

"Let's wait and see," Huang Zhenxin, head of enterprise business at Moonshot AI, said yesterday, adding that the parameter scale of Kimi K3 will continue to expand in the future.

Yesterday marked the Beijing-based firm’s first public comments since launching K3 on July 16. The latest version of its Kimi AI model has a total of 2.8 trillion parameters and a context window of one million tokens. In tests, its overall intelligence level ranked just behind Claude Fable 5 and GPT-5.6 Sol, which are widely considered to be the world’s two leading AI models.

K3 became so popular after its launch that Moonshot AI had to suspend new member registrations on July 19. Musk said the model is "impressive," and several foreign media outlets and research institutions labeled its release as a "DeepSeek 2.0 moment," referring to the global impact made by the Chinese AI firm with its releases in January last year.

Huang stated that the performance leap of K3 is based on three proprietary technologies. First, its MoonClip second-order optimizer allows 20 trillion tokens of training data to deliver performance comparable to 40 trillion tokens, while halving training costs and computational power consumption for an equivalent performance.

Second, the Kimi Linear Tension mechanism addresses performance degradation in ultra-long tasks, expanding the context window 10-fold. And third, its Attention Residuals technology improves information transfer between multi-layer networks, enabling the stable training of the 2.8 trillion-parameter model while boosting inference efficiency by 25 percent.

K3 marks a new milestone for Chinese AI models in achieving pricing power, according to a recent report by Goldman Sachs. Morgan Stanley commented that K3 shows Chinese AI models are achieving a "comprehensive catch-up" with leading US models in terms of scale, performance, and pricing.

In terms of commercialization, Huang revealed that 70 percent of the firm’s revenue comes from application programming interface services, mainly driven by data platform clients. The long-term focus will be on high-difficulty productivity scenarios such as programming, finance, law, and scientific research, he said, adding that the company is not planning to venture into the entertainment, text-to-image, or text-to-video domains.

According to media reports, Moonshot AI’s annual recurring revenue has surpassed USD300 million as of the middle of last month.

Editor: Tom Litting