Moonshot pauses Kimi K3 signups as demand outstrips GPUs
Moonshot AI halted new Kimi K3 subscriptions after demand surged sixfold, as third-party tests put the open-weight model within reach of Claude Fable 5.
Moonshot AI has stopped accepting new subscriptions for Kimi K3, saying demand in the 48 hours after launch pushed its GPU fleet to the edge of capacity — a scarcity signal that arrived just as independent benchmarks placed the Chinese model alongside the strongest US frontier systems.
Too popular to serve
Kimi K3 launched on July 16 as Moonshot's most ambitious model: a 2.8-trillion-parameter mixture-of-experts system with the company's Kimi Delta Attention design, a one-million-token context window, and native vision — the largest model on an open-weight track to date, well beyond DeepSeek V4's 1.6 trillion parameters. Demand reportedly surged roughly sixfold within days. On July 19 the company said it would pause new signups to protect existing subscribers' experience and reopen access in batches; full weights are expected to be published around July 27, which could offload inference demand to third-party hosts. The crunch comes as Moonshot reportedly pursues a Hong Kong IPO at a valuation near $30 billion.
Benchmarks close the gap
The demand has empirical backing. Inference provider Fireworks AI published an evaluation across roughly 1,030 agentic tasks concluding K3 performs at near parity with Anthropic's Claude Fable 5 — 92.4% versus 92.6% on its software-engineering suite — with K3 stronger in symbolic math and terminal work, Fable ahead on visualization and language breadth. Fireworks argues the interesting result is routing: an oracle router mixing both models hit 93% while cutting costs dramatically, with K3 economically handling most traffic. The launch's shockwave was macro as well — the release contributed to a slide of more than 20% in semiconductor stocks from June peaks and fed directly into this week's US sanctions debate.
Why it matters
An open-weight model turning away paying customers for lack of GPUs is the clearest demand signal Chinese frontier AI has yet produced. If the July 27 weights release lets anyone serve near-Fable capability on commodity infrastructure, the pricing power of closed frontier models faces its most direct test to date — precisely the scenario driving both Washington's alarm and Jensen Huang's counterargument.