Image: Euronews / euronews.com Kimi K3 Topped the Coding Leaderboard. Two Days Later, Moonshot Ran Out of Servers.
Moonshot AI paused new Kimi K3 subscriptions on July 19 after demand pushed its GPU infrastructure close to the limit within 48 hours of launch.
Moonshot AI launched Kimi K3 on July 16, watched it climb to the top of Arena’s front-end coding leaderboard, and then suspended new subscriptions three days later. The demand was real. The servers weren’t ready for it.
In a post on X, Moonshot said demand had “pushed close to the limits” of its infrastructure within 48 hours of launch. Existing subscribers keep access. New sign-ups are paused while the company adds capacity, and spots will reopen gradually “in batches.”
What broke
Kimi K3 is a 2.8 trillion-parameter mixture-of-experts model — the largest open-weight model released so far. Its compute profile is unusual even by MoE standards: the context window runs up to one million tokens, and the model supports native vision. That combination puts a heavy load on inference hardware.
Omdia analyst Lian Jye Su told Euronews that Moonshot likely underestimated K3’s popularity, and that its unusual compute demands make capacity allocation expensive to scale quickly. That’s not a new pattern for Chinese AI labs — it’s a version of what happened when DeepSeek R1 went viral in early 2025 and saturated its API.
The suspension happened fast because the model was good enough that people actually wanted to use it. On Arena’s front-end coding benchmark, K3 scored 1,679 points and ranked first, ahead of Claude Fable 5 and every other model in blind developer voting. That kind of result drives real adoption.
What changes
Moonshot said it plans to split its subscription tiers into separate Kimi and Kimi Code memberships once capacity stabilizes. Existing subscribers won’t be downgraded or charged more during the pause.
The open weights, originally slated for July 27, are presumably still on schedule. Running K3 locally would sidestep Moonshot’s infrastructure entirely, though the hardware requirements for a 2.8T model are substantial.
The API remains accessible. Users accessing Kimi K3 through the Moonshot API or via Kimi Code are not affected by the subscription pause.
Sources: Euronews — ABC News — Dataconomy