Beijing-based Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open-weight model that debuted at #1 on Arena.AI’s Frontend Code leaderboard with 1,679 points, and by Friday’s close global markets had translated the benchmark into a semis selloff. TSMC fell 7% on the same day it posted a 77% jump in quarterly operating profit. SoftBank dropped 9%. Chinese rival Z.ai plunged nearly 30% in Hong Kong. The Nasdaq 100 slid 1%, Nvidia gave up 1.2%, and Meta fell more than 2.4%.
The reflex is familiar. “an over-reaction shockingly similar the DeepSeek panic,” Patrick Moorhead of Moor Insights and Strategy told CNBC, and the January 2025 DeepSeek R1 selloff is the obvious template, right down to the Friday timing and the reflex to sell TSMC and Nvidia against earnings that hadn’t yet been printed.
What’s different this time is architecture and price. K3 activates 16 of 896 experts per token, ships a 1-million-token context window, and introduces Kimi Delta Attention, a hybrid linear-attention scheme, alongside a feature Moonshot calls Attention Residuals. The company claims roughly 2.5x scaling-efficiency gains over Kimi K2. Kernel benchmarks were run on Nvidia H200s; MiniTriton results were charted on the export-restricted L20. Full weights land July 27.
Pricing is the sharper knife. K3 lists at $15 per million output tokens against Claude Fable 5’s $50, with cache-hit input at $0.30 and uncached input at $3. That uncached number is roughly five times what Kimi K2 charged a year ago, which is its own tell: Moonshot is pricing into demand, not undercutting it. Bank of America analysts led by Alex Liu wrote that “Despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models.”
Enterprise adoption is already visible. Cursor Composer 2 ran atop Kimi 2.5. DoorDash CTO Andy Fang described routing “lower-level work to Kimi K2.6.” Thinking Machines shipped Inkling into the same competitive frame. Moonshot raised $2 billion in May at a valuation above $20 billion on annual recurring revenue exceeding $200 million.
The complication is that every headline claim still needs independent verification, and the priors aren’t clean. Anthropic has accused Moonshot of using 3.4 million Claude exchanges to train its models, an allegation that sits underneath the Fable 5 and Opus 4.8 comparisons Moonshot invites, and next to the GPT 5.6 Sol and 5.5 benchmarks the release positions against. The July 27 weights drop is when the arena stops being a leaderboard and starts being a lab.
Markets priced the story before the story could be checked. That, more than the parameter count, is the structural fact.
Sources
- 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://fortune.com/2026/07/16/moonshots-kimi-k3-pushes-chinese-ai-into-fable-level-territory/
- https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3