Moonshot AI’s Kimi K3 Sparks a New DeepSeek Moment

Moonshot AI’s Kimi K3 has climbed to the top of a major AI coding leaderboard, prompting comparisons with DeepSeek’s 2025 breakthrough and renewed debate over China’s ability to challenge leading U.S. labs.

The Beijing-based startup released the model Friday. Kimi K3 is one of several Chinese large language models gaining global attention for their lower costs and customizable source code. Large language models power chatbots and other AI tools by processing vast amounts of digital data.

Shortly after launch, Kimi K3 topped the AI coding leaderboard operated by Arena, a platform created by researchers at the University of California, Berkeley. Arena also ranked it ninth worldwide for text queries.

“Kimi K3 seems really good, closest to the frontier yet.”

Mollick added that the model “cannot write a good murder mystery,” although he said other models also struggle with that task. “That remains the jaggedest of frontiers” in AI development, he wrote.

Recommended reading

The philosopher shaping DeepMind’s AI ethics

Tech writer and investor Kevin Xu predicted a market reaction similar to the “DeepSeek moment,” referring to the Chinese startup’s 2025 model release, which challenged assumptions about U.S. dominance in AI.

Kimi K3's model size and performance

Moonshot AI calls Kimi K3 the world’s first open-source model of its size. It contains around 2.8 trillion parameters, the internal variables that help models handle complex requests. Anthropic and OpenAI have not disclosed the parameter counts of their leading models.

The company said Kimi K3 delivered “frontier-level performance” across its evaluation suite and consistently outperformed the other models tested. It also acknowledged that the model’s overall performance still trails the most powerful proprietary systems from Anthropic and OpenAI.

Hussein Abbass, a computing professor at UNSW Canberra, said Kimi K3 appears strong at coding but that its performance across the full range of foundation-model tasks remains unknown.

“I wouldn’t say they need to be worried. But they shouldn’t stand still.”

Abbass said AI performance depends not only on the model, but also on the hardware, data centers and supply chains needed to run it.

China and U.S. AI competition

Kimi K3 follows other heavily promoted Chinese models, including Zhipu AI’s GLM-5.2. Its release came during a major technology conference in Shanghai, where President Xi Jinping called for international cooperation on AI governance Friday.

The United States restricts exports of powerful microchips used to train and run AI systems in China. Earlier this year, the U.S. said it was approximately eight months ahead of China in the strategic field.

At the same time, the Trump administration delayed the public release of top-end models from Anthropic and OpenAI over concerns that they could help hackers break into online systems. Mollick questioned whether the growing performance of open-weights models would affect how quickly those companies are permitted to release new systems.

The source credits Andrew Zinin, who holds a master’s degree in physics and has research experience.