Moonshot's 2.8T Kimi K3 challenges US frontier labs
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Moonshot AI released Kimi K3, a 2.8T-parameter model built on its Kimi Delta Attention with 1M-token context, and plans to publish full model weights by July 27, per the company's blog
- Kimi K3 identified itself as Anthropic's Claude in at least one conversation, according to Wccftech, which framed the slip as 'betraying its distilled origins'
- Gavin Baker (@gavinsbaker) argued that competing open models like Kimi K3, Grok 4.5, and Muse 1.1 may prevent 2-3 frontier labs running 90% inference margins from squeezing other layers of the AI ecosystem
- The Decoder flagged that Kimi's open release 'signals the end of super-cheap Chinese AI,' undercutting a pricing expectation that had defined the early-2026 open-weights segment
- Bloomberg reported the model 'surprised investors, fueling a tech rout,' and ZeroHedge called it the latest 'DeepSeek moment' that sparked a 'chip meltdown' in futures
- Axios framed the launch as 'China just erased America's AI lead,' while WSJ and International Business Times echoed the line that 'Silicon Valley [is] no longer clearly leading'
Why it matters: Frontier-class open weights arriving on a hard July 27 deadline reset the negotiating position of every API customer against the US labs running the 90% inference margins Baker flags; if Kimi K3 benchmarks hold once weights land, Moonshot — not Anthropic or OpenAI — becomes the price-setter for the open tier, and the inference-margin floor those frontier labs have priced in collapses.
