On July 27, 2026, Moonshot AI published the full weights of Kimi K3 on Hugging Face, the biggest open-weight release in AI history. The mixture-of-experts model has 2.8 trillion parameters, a 1 million token context window, and weighs about 1.4 TB in 4-bit MXFP4 (roughly 5.6 TB at 16-bit). Running it takes a multi-GPU cluster of 80 GB cards, and the license terms were not published ahead of the release.
On July 22, 2026, White House OSTP Director Michael Kratsios stated that Moonshot AI built Kimi K3, a 2.8 trillion parameter model, by running large-scale distillation against Anthropic's Fable 5 and tapped export-restricted Nvidia GB300 chips through servers in Thailand. Anthropic earlier reported over 3.4 million Claude exchanges through fraudulent accounts, but experts note Fable 5 was public only 15 days before K3 shipped, and Moonshot denies wrongdoing. Treasury has floated sanctions.
Beijing startup Moonshot AI paused all new Kimi K3 sign-ups on July 19, 2026, 48 hours after launch, after demand pushed its GPU clusters to full capacity; existing subscribers keep access. K3 is a 2.8 trillion parameter mixture-of-experts model that has beaten GPT-5.6 Sol and Claude Fable 5 on several benchmarks, though users note it runs slower than top US models. Moonshot says it will release the model weights on July 27, making K3 the largest open-weight frontier model in the world.
Moonshot AI, the Beijing company behind the Kimi chatbot, has told investors it is restructuring for a Hong Kong IPO (per SCMP and The Next Web), dismantling its offshore VIE structure and merging into a single listing vehicle, a step Beijing effectively required. Chinese tech media report the listing could complete within six months, with analysts expecting late 2026 or early 2027; Moonshot's valuation sits around $20 billion after raising about $2 billion this year. The move lands days after Kimi K3 hit number one on Arena's Frontend Code Arena, with open weights promised by July 27.
On July 16, 2026, Moonshot AI announced Kimi K3, a 2.8 trillion parameter Mixture-of-Experts model with a 1 million token context window that debuted at number one on Arena.ai's Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). On July 17 the Nasdaq dropped 1.4 percent, Nvidia slipped 2 percent, and the Philadelphia Semiconductor Index extended its slide into bear-market territory in what analysts called a DeepSeek flashback. The weights are not out yet: Moonshot says it will publish them by July 27.