Moonshot AI, a Beijing-based startup, unveiled its Kimi K3 large language model, which promises performance comparable to Anthropic's Fable 5 and OpenAI's GPT-5.6, but at a significantly lower cost. Kimi K3 boasts 2.8 trillion parameters, making it the largest open model released to date, and will be available for free download from July 27. This release has caused significant market disruption, with the Philadelphia Semiconductor Index experiencing its worst week in over 15 months, falling 12.5%. Nvidia, AMD, and Broadcom all saw sharp declines, though the index remains up over 60% for the year.
The market reaction mirrors previous shocks, such as when China's DeepSeek caused Nvidia to lose $589 billion in one day in January 2025. Analysts noted the Kimi K3's coding performance, competitive pricing, and open-weight release as factors closing the gap on US AI labs. Moonshot charges $3 per million input tokens, while Anthropic's Fable 5 costs $10 per million input tokens, and OpenAI charges $5 for GPT-5.6 Sol. The Kimi K3 also topped Arena's coding leaderboard with 1,679 points, surpassing Claude Fable 5.
The introduction of Kimi K3 has reignited discussions about the AI bubble and the American AI lead. While US models still hold top scores on most major benchmarks, the free availability of Kimi K3 from July 27 could shift the competitive landscape. US chip stocks bore the brunt of the impact, with Chinese AI stocks like Zhipu and MiniMax also seeing declines of 28% and 16% respectively. The ability for Moonshot to train its recent models on Nvidia's export-grade H800 chips, despite Washington's efforts to restrict chip exports, has also raised eyebrows.
