Kimi K3 is a huge open-weight AI model released by the Chinese startup Moonshot AI in July 2026. It matters beyond the usual model-launch news cycle because its release briefly rattled global markets, dragging chip stocks into a sell-off and helping Apple momentarily overtake Nvidia as the world’s most valuable company. Here’s what the model actually is, who built it, and why an AI release moved stock prices.

What Is Kimi K3?

Kimi K3 is a mixture-of-experts (MoE) language model with 2.8 trillion total parameters — Moonshot calls it the largest open-weight model released to date. The MoE design means the model doesn’t use all of its parameters for every request: K3 activates only 16 of its 896 “expert” sub-networks per token, roughly 1.8% of the total, which keeps the computing cost of each answer far lower than a dense model of the same size would require. Our explainer on mixture-of-experts models covers this architecture in more depth.

The model also supports a 1-million-token context window, can process images natively, and introduces two new efficiency techniques Moonshot calls Kimi Delta Attention and Attention Residuals, aimed at cutting compute costs on long conversations without hurting reasoning quality. In Moonshot’s own benchmark suite, K3 trails Anthropic’s newest flagship model and OpenAI’s latest reasoning model on overall capability, but it topped independent, blind evaluations of coding tasks, including first place in the Frontend Code Arena test, ahead of larger and more expensive rivals.

Who Is Moonshot AI?

Moonshot AI is a Beijing startup founded in March 2023 by Yang Zhilin, Zhou Xinyu, and Wu Yuxin, three researchers who met at Tsinghua University. Yang, the CEO, holds a PhD from Carnegie Mellon and co-authored two influential research papers, Transformer-XL and XLNet, before returning to China to build foundation models. The company is backed by a $1 billion investment from Alibaba and is counted among China’s so-called “AI Tigers,” a group of startups racing to build large language models domestically. Its consumer product, the Kimi chatbot, launched in October 2023 and now serves tens of millions of monthly users; K3 is the model now powering it.

Why It Rattled the Stock Market

The market reaction wasn’t about the model’s benchmark scores — it was about what an efficient, freely downloadable model implies for spending on US computing infrastructure. Much of the current AI boom rests on the assumption that ever-larger models require ever-more Nvidia chips, justifying hundreds of billions of dollars in planned data-center spending. A model that reaches near-frontier performance while activating under 2% of its parameters per query undercuts that assumption, at least in investors’ eyes, by suggesting the same results might be reachable with far less hardware.

The pattern echoed DeepSeek, the Chinese lab whose efficient model rattled markets in a similar way in early 2025. In the days after K3’s release, the Philadelphia Semiconductor Index fell into bear-market territory, and Apple briefly passed Nvidia in market capitalization for the first time since April 2025, as investors rotated toward companies seen as less exposed to a potential slowdown in AI infrastructure spending. Nvidia reclaimed the top spot by the end of the trading day, but the episode illustrated how sensitive chip valuations have become to any evidence that AI capability doesn’t require unlimited compute.

How to Try Kimi K3

Anyone can chat with Kimi K3 for free through Moonshot’s own app at kimi.com. Developers can call the model through Moonshot’s API, and the full model weights are due for release on July 27, 2026, after which anyone with sufficient hardware will be able to download and run the model directly, as Moonshot has done with earlier Kimi releases.

In the news

For more on the market fallout from K3’s release, see our report on how Apple briefly overtook Nvidia as chip investors reassessed AI spending.