Kimi K3: China’s New AI Sensation

The success appears to have come faster than expected: Moonshot AI has temporarily suspended new registrations for its Kimi K3 AI model because its existing server infrastructure cannot keep up with the surge in users. So what is behind the hype, and what exactly is inside the model?

Just 48 hours after Kimi K3 was released, Moonshot AI announced on X that its systems were nearing capacity. According to the company, existing users can continue to access the model as usual. For new users, however, registration will remain closed for the time being.

Moonshot said it plans to gradually expand capacity and then enable new accounts in batches. The company did not say how long the process would take.

Capacity bottlenecks shortly after a model launch are currently not unusual among Chinese AI companies. New releases repeatedly attract so much international attention that providers’ technical infrastructure struggles to handle the influx. At the same time, the incident highlights the increasingly intense competition between Chinese and US AI developers.

How Good Is Kimi K3?

Kimi K3 is part of a growing lineup of affordable, open Chinese AI models that have recently gained international attention. The trend began in early 2025 with DeepSeek’s original model, which triggered significant market reactions, followed by the successor V4. Both releases helped establish China as an increasingly serious player in global AI development.

One common feature of many leading Chinese AI models is the public release of their model weights. Unlike purely cloud based services, this allows developers to inspect the models’ internal architecture and adapt them for their own use cases.

Moonshot launched Kimi K3 last week with 2.8 trillion parameters, which the company describes as the largest open weight model released to date. According to reports from the South China Morning Post, the model is said to have outperformed systems such as GPT-5.6 Sol and Claude Fable 5 in several benchmark tests. In the Arena rankings for frontend coding tasks, Kimi K3 reportedly took the top spot following its release.

The model’s strengths include a one million token context window and a strong focus on complex programming tasks and agentic workflows. These include independently navigating large code repositories, using tools, and debugging based on images, log files, and test results. Kimi K3 can also process images and other data formats in addition to text, making it potentially useful for applications in game development and 3D design.

Meet Kimi K3

The Real Reason Behind the Hype: Price

The surge in demand is largely driven by a significant price gap compared with established US providers. Based on published API pricing, Kimi K3 costs considerably less per million tokens than offerings from OpenAI and Anthropic.

The figures are based on published API list prices as of July 2026. As experience shows, these prices can change frequently. At comparable usage levels, Claude Fable 5 is therefore around 3.3 times more expensive than Kimi K3.

Moonshot also plans to release the full model weights by July 27, 2026. This could allow companies to run Kimi K3 on their own hardware rather than relying on Moonshot’s cloud infrastructure. For organizations dealing with sensitive data in particular, that could be an additional selling point.

Impact on US Technology Stocks

The announcement also had an impact on financial markets. Shares of major US technology companies came under pressure as investors worried that cheaper Chinese AI models could increasingly put pressure on the pricing strategies of American providers. This comes despite US export restrictions that continue to limit Chinese companies’ access to the most powerful chips.

(lb)