Moonshot pauses new Kimi K3 subscriptions as demand outruns GPU capacity

Moonshot AI has paused new subscriptions to Kimi K3 after demand pushed close to the limits of its current GPU capacity, the company said via its official Kimi account. Moonshot launched the flagship model July 16 and hit that ceiling within about 48 hours. Existing subscribers are unaffected.

The Kimi K3 subscription pause is a capacity problem, not a product retreat. Kimi K3 is a 2.8-trillion-parameter open-weight mixture-of-experts model with a context window of up to 1 million tokens, and interest in it outran the compute Moonshot had provisioned almost immediately after release. The company said it is expanding GPU capacity and will reopen new subscription slots in controlled batches rather than all at once.

Moonshot also said it will split its membership into two tiers to match compute to how people use the model: a Kimi Membership covering the Kimi Web, App and Work products, and a separate Kimi Code Membership aimed at coding workflows, which tend to consume far more tokens per session.

The episode is a familiar one for open-weight labs shipping frontier-scale models: the cheaper and more capable the model, the faster demand saturates whatever inference capacity sits behind it. Moonshot has said it will publish the full open weights for Kimi K3 on July 27, which would let outside developers run the model on their own hardware and take some load off the company’s hosted service.

Moonshot has not said how much capacity it is adding or how long the pause will last, and the “close to the limits” framing is the company’s own characterization rather than a disclosed utilization figure. Whether the batched reopening clears the backlog within days, or the constraint drags into the open-weight release, will show how badly Kimi K3 is compute-bound.