Moonshot pauses new Kimi subscriptions three days after K3 launch, citing demand

The numbers are drawing attention anyway. Tom's Hardware reported that Arena's blind developer testing ranked K3 first in its Frontend Code evaluation, ahead of Claude Fable 5.6 Those standings are claims until outsiders can reproduce them, and the weights that would allow that are not out yet.

Why it matters. A subscription pause three days into a launch signals real, paying demand for a Chinese open-weight flagship. It also shows where that demand hits a wall: compute. U.S. export controls keep the most capable AI chips fenced off from Chinese buyers, even as Washington opens license-free chip channels for partners it trusts, as it did for approved buyers in the UAE this month.

For a lab that cannot simply buy its way out of a shortage, the membership split reads as compute rationing built into the product: metering scarce GPU time by workload, with coding traffic on one plan and everything else on another.

The open-weights race is compressing. K3 was announced one day after Thinking Machines released Inkling, a 975-billion-parameter model whose weights shipped immediately under Apache 2.0 — two frontier-scale open releases in two days, from labs on opposite sides of the export-control line. By Moonshot's own count, Kimi models have set the upper bound of open-model sizes for nine of the past twelve months.

Moonshot was founded in March 2023 by Yang Zhilin and two Tsinghua University schoolfriends, counts Alibaba and Tencent among its backers, and is one of the six Chinese labs known as the "AI Tigers." It has been squeezed by its own popularity before: a surge of users knocked the Kimi chatbot offline for two days in March 2024, forcing a public apology.7

What we still don't know. Moonshot published no numbers: no subscriber counts, no wait-list size, and no reopening date beyond the promise of batches. It has not said what hardware serves subscribers in production or where added capacity will come from — though its technical blog mentions Nvidia H200s and an unnamed alternative GPU vendor in engineering examples, and recommends clusters of 64 or more accelerators to run the model well.8 Whether the July 27 weights date holds is now the next test of how open K3 actually is.