Moonshot AI, the Alibaba-backed Beijing startup, released Kimi K3 on July 16, 2026. The 2.8-trillion-parameter model is billed as the largest open-weight AI model ever released to the public. A day later, TSMC shares fell 7 percent despite a 77 percent jump in quarterly profit, SoftBank dropped 9 percent, and Chinese AI stock Z.ai plunged nearly 30 percent in Hong Kong trading. Nvidia briefly lost its title as the world's most valuable company to Apple.

Kimi K3 uses a sparse mixture-of-experts architecture: only a small share of its total parameters activate for any given request, which keeps inference costs down despite the model's scale. It supports a 1-million-token context window and comes in two versions: K3 Max for chat and agentic tasks, and K3 Swarm Max for large-scale parallel processing. Two new architectural features, Kimi Delta Attention and Attention Residuals, let K3 use 21 percent fewer output tokens than Kimi K2.6 on comparable tasks.

Why did the stock market react so hard?

Kimi K3 upends investors' assumption that the US AI edge rests on far greater compute spending. Moonshot says it reached near-frontier performance without access to the top-tier Nvidia chips that US export controls bar from China, relying on architectural efficiency instead of raw computing power.

"The AI ecosystem in China is probably much better than people thought," said Paul Triolo, a partner at DGA-Albright Stonebridge Group. Lian Jye Su, chief analyst at Omdia, added that models like this "can be run at a fraction of the cost that OpenAI charges its clients." Not every analyst thinks the shock will last. Patrick Moorhead, CEO and chief analyst at Moor Insights and Strategy, called it "an over-reaction shockingly similar to the DeepSeek panic," referring to the market turmoil DeepSeek triggered in early 2025.

The benchmark scores and pricing behind the story

Artificial Analysis put Kimi K3's Elo score at 1,547, up 732 points from Kimi K2.6. That is still below Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol overall, but Kimi K3 beats Claude Opus 4.8 and GPT-5.5 on several coding and agentic tests, and it tops the Arena.ai frontend-code benchmark ahead of Claude Fable 5. The pattern echoes MiMo Code, another Chinese coding model that claimed to rival Claude Code.

Moonshot priced the Kimi K3 API at $3 per million input tokens and $15 per million output tokens, the highest of any Chinese AI lab, but still about half the per-task cost of Claude Opus 4.8 and roughly in line with Claude Sonnet pricing. That marks a shift from the cheap-pricing strategy DeepSeek and Qwen have used to compete: Moonshot is charging a premium because its performance is close to frontier level.

Bank of America analysts said Chinese AI companies have shown they can make major progress by improving how they train and design models even with restricted chip access, and Kimi K3's arrival is likely to sharpen the debate in Washington over how effective chip export controls really are.

Who is behind Moonshot AI?

Moonshot AI was founded in March 2023 by Yang Zhilin, now 34, along with Zhou Xinyu and Wu Yuxin, three schoolmates from Tsinghua University. Yang finished his PhD at Carnegie Mellon University in just four years. "Not only did he complete his PhD in just four years, but he also made truly fundamental contributions to machine learning," said Russ Salakhutdinov, a computer science professor there, of his former student.

Alibaba led a $1 billion funding round in February 2024 at a $2.5 billion valuation, putting in about $800 million for a 36 percent stake in Moonshot. Tencent, Meituan, HongShan, IDG Capital and 5Y Capital also invested. In May 2026, Moonshot raised about $2 billion in new funding at a valuation above $20 billion, as it weighs a Hong Kong IPO under consideration since early this year.

Moonshot's release pattern has scaled up every six months: from the original Kimi's 200,000-character capacity in October 2023, to the open-source, 1-trillion-parameter Kimi K2 in July 2025, to Kimi K2 Thinking in November 2025, and now K3's 2.8 trillion parameters a year later. The release also comes shortly after Anthropic's Fable 5 and Mythos 5 were shut down three days after launch and then restored worldwide, an episode that shows how fast the frontier-model race is moving on both sides of the Pacific.

What to watch

Kimi K3's full weights are due out on July 27, 2026. That release will show how easily developers outside China, including in Indonesia, where many currently pay for Claude or GPT API access in US dollars, can actually download and run the model on their own. How tech stocks move in the following weeks will reveal whether the July 17 shock holds up like the DeepSeek fallout did in early 2025, or fades once independent benchmarks retest Moonshot's claims.