Moonshot AI says Kimi K3 narrows China’s gap with U.S. AI leaders

Moonshot AI, a Beijing startup founded in 2023, released a new artificial intelligence model Friday that it says can compete more closely with top systems from OpenAI and Anthropic. For investors watching the AI trade, the reaction was immediate: shares of several Chinese AI rivals fell as the market reassessed who may be gaining ground.

The new model, called Kimi K3, is China’s largest AI model to date, according to Moonshot. The company said it has 2.8 trillion parameters, a measure of the size of the neural network. In plain terms, parameters are the internal settings a model learns during training, and larger models often have more capacity to recognize patterns and handle complex tasks.

Moonshot said Kimi K3 still ranks behind Anthropic’s Claude Fable 5 and OpenAI’s GPT 5.6 Sol when looking at overall performance. But the company said its model beat other tested systems, including Anthropic’s Claude Opus 4.8 and OpenAI’s GPT 5.5, on benchmarks covering areas such as coding and general agent tasks.

Benchmarks are standardized tests used to compare AI models. They are useful for investors because they help show whether a company’s model is improving, though they do not capture every real-world use case.

China’s AI race gets a new marker

The launch comes as U.S. and Chinese companies compete to build more capable AI systems. Chinese models have been gaining users among Western companies as their performance improves and their usage costs remain below those of the most advanced U.S. lab models, CNBC has reported. CNBC has also reported that U.S. lawmakers are weighing ways to limit adoption of Chinese AI models by American companies.

Bank of America analysts led by Alex Liu said Kimi K3 shows that Chinese labs can keep making progress even with limits on computing power. “Despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models,” the analysts wrote in a note.

Pre-training is the early, resource-heavy stage when an AI model learns from large amounts of data before it is tuned for specific tasks. Compute refers to the chips and data-center capacity used to train and run these systems.

Liu said K3 lifts expectations for the rest of China’s AI sector. “K3 raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs,” he wrote.

Rival stocks take a hit

The market response was sharp among Moonshot’s Chinese competitors. Z.ai, which released a new model in June, dropped 28% Friday, according to CNBC. MiniMax Group, another Chinese model developer, fell 16%.

Alibaba also moved lower. The company, which develops the Qwen family of AI models, had gained earlier in the week after news that it was partnering with Apple in China, CNBC reported. Alibaba shares fell 4% Friday.

Bank of America’s Liu said Alibaba may still benefit from broader demand for AI training and usage through its cloud business because computing resources remain tight. But he added that the “open-source leader” story around Alibaba’s Qwen models may face pressure after Moonshot’s release.

For retail investors, the takeaway is less about one model test and more about how quickly AI leadership can shift. Model releases are becoming market-moving events because they can affect perceptions of cloud demand, pricing power and which companies are closest to the frontier.

This story draws on original reporting from CNBC.