Startups

Moonshot's Kimi K3 goes open-weight, and the enterprise AI stack braces for impact

Moonshot AI's 2.8-trillion-parameter model is now downloadable, sending Chinese rivals' shares tumbling and reigniting Washington's fight over open-source AI from China.

By Giulia Donatello · Startups & funding · 2026-07-29 · 3 min read

Photo: Unsplash / Alexandre Debiève — Close-up of a green circuit board with silver traces, representing AI compute infrastructure.

Beijing-based Moonshot AI has made the full weights of Kimi K3, a 2.8-trillion-parameter model with a one-million-token context window, publicly downloadable, and the reaction across the enterprise AI stack has been immediate. Z.ai fell 28% after the launch. MiniMax Group dropped 16%. Two days in, Moonshot had to suspend new subscriptions because demand overwhelmed capacity.

The company claims K3 outperforms every rival except Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 on overall capability, and the business numbers give the claim teeth. Daily revenue has grown at least six-fold since launch. Annual recurring revenue jumped from $200 million in April to $300 million in June. Moonshot, backed by Alibaba and Tencent, raised $2 billion in May at a valuation north of $20 billion and is now targeting $50 billion in a fresh round, with a possible Hong Kong IPO as soon as this year.

The strategic move isn’t the parameter count. It’s the licensing. Chelsey Tam, senior equity analyst at Morningstar, notes that many organizations value the ability to self-host open models, retaining the weights and control over sensitive data rather than depending on closed-model providers. That’s the pitch every hyperscaler distribution channel, AWS Bedrock, Azure Foundry, Google Vertex AI, will now have to answer for.

Washington heard it too. Michael Kratsios, White House OSTP Director, has alleged that Moonshot trained K3 on banned Nvidia chips and ran large-scale distillation against American models. Moonshot hasn’t responded to requests for comment. OpenAI and Anthropic have reportedly lobbied regulators over open Chinese models, though even inside OpenAI the story is complicated: Dean Ball, its head of strategic futures, called K3 “a very good model” whose performance probably can’t be “explained away by distillation or anything like that.”

Patrick Moorhead of Moor Insights and Strategy called the market reaction “an over-reaction shockingly similar the DeepSeek panic.” The reference is doing work. DeepSeek in early 2025 also produced a share-price convulsion that faded within weeks while the underlying dynamic, cheap capable Chinese open weights sitting one download away from any enterprise procurement team, didn’t.

Sources