Moonshot AI released Kimi K3 late Thursday, a 2.8-trillion-parameter open-weight model that’s now the largest ever shipped with weights in the open, and by Friday afternoon the trade it threatened had already reordered the tape. TSMC fell 7%. SoftBank dropped 9%. Nvidia slid 1.2%, Meta more than 2.4%, and by 2 p.m. Eastern the Nasdaq 100 was off 1.0%. The Beijing lab, backed by Alibaba and Tencent and reportedly seeking $2 billion at a roughly $30 billion valuation, said K3 performs “competitively” against Anthropic’s Claude Fable 5.
The parameter count is the headline, but the framing is the story. K3 is nearly double the size of DeepSeek V4 Pro’s 1.6 trillion, and it ships open. That combination is what the selloff is actually pricing: not a single benchmark result, but a second data point in a pattern where Chinese labs, operating under export controls, keep producing frontier-adjacent systems and then giving the weights away.
Bank of America analysts led by Alex Liu put it plainly: “despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains.” Translated out of sell-side register, the compute-moat thesis underwriting a lot of 2025’s capex just got another stress test.
Not everyone thinks the tape is reading it correctly. Patrick Moorhead, chief analyst at Moor Insights and Strategy, posted on X that the reaction looked like “an over-reaction shockingly similar the DeepSeek panic” of January 2025, when the same names sold off on the same theory before recovering. TSMC’s quarterly operating profit is still up 77%. The fundamentals didn’t move on Friday; the vibes did.
The China-listed AI names took the harder hit. Zhipu fell 27.7% in Hong Kong, Z.ai roughly 30%, MiniMax 16.5%: a rotation inside the Chinese AI complex toward whoever shipped the biggest open model this week, not a broad vote of confidence. Meanwhile the coding evaluations, Program Bench and SWE Marathon, along with third-party readouts from Arena.ai, Vals AI, and Artificial Analysis, put K3 in range of GPT-5.5, GPT-5.6 Sol, Claude Opus 4.8, Fable, and Mythos. Simon Koser, chief product officer at Tzafon, called K3’s coding “legitimately impressive” and added the sentence that actually matters for downstream deployment: “Cost has become a huge thing for some of these labs.” Cursor already integrated Kimi 2.5 into its Composer 2 agent. K3 will follow the same path.
That’s the structural read. Every open-weight release from a Chinese lab compresses the price American incumbents can charge for equivalent capability, and every compressed price makes the trillion-dollar capex forecasts harder to underwrite. The market spent Friday relearning something it learned in January and then forgot.
Sources
- https://finance.yahoo.com/technology/ai/articles/chinas-moonshot-unveils-worlds-largest-020622030.html
- https://www.bloomberg.com/news/articles/2026-07-17/china-s-powerful-new-moonshot-ai-model-closes-gap-with-us-rivals
- https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
- https://fortune.com/2026/07/17/china-moonshot-kimi-k3-markets-china-ai/
- https://www.scmp.com/tech/tech-trends/article/3360885/moonshot-ai-unveils-worlds-largest-open-source-ai-model-china-narrows-gap-us-rivals