Kimi K3 becomes world's largest open model, forces subscription halt within 48 hours

Kimi K3 Became the World's Largest Open Model. Then It Had to Stop Selling.

Moonshot AI released Kimi K3 last week, a 2.8 trillion-parameter system and the world's largest open-weights AI model. Forty-eight hours later, the Beijing-based AI startup stopped accepting new subscriptions because demand overwhelmed its servers 1.

That freeze, not the parameter count, is the detail worth scrutinizing. Frontier model launches always trigger traffic spikes. But Moonshot didn't see a bump. Demand pushed "close to the limits of our current capacity" within 48 hours, the company wrote on X on July 19 1. Moonshot said it would reopen subscriptions in batches as it adds compute 1.

The reason developers piled in so fast: K3 delivers. By the measure of front-end coding capability, it topped the rankings on Arena, an AI model evaluation platform, beating every major model in that category right after release 1. Developers weren't window-shopping. They loaded the model hard enough to buckle the infrastructure serving it.

That buckle is the bottleneck the open-weights movement has been building toward. Open models have spent more than a year closing the capability gap with closed systems from OpenAI and Anthropic. DeepSeek's V4 continued a trajectory that began when DeepSeek shook global markets in early 2025 and established Chinese open-source models as credible competitors 1. The models are now good enough, and cheap enough, that serving them at scale has become the hard part.

Lian Jye Su, a chief analyst at the technology research and advisory group Omdia, said the halt shows Moonshot AI "does not have sufficient compute chips to serve the current surge in demand" 1. He noted that K3 is "very demanding" in compute requirements, making allocation expensive and difficult 1.

For developers picking a stack, K3's freeze exposes a trade-off that's getting harder to ignore. The model is open-weights: anyone can download it, inspect it, and build on it. But running a model this size in production requires serious hardware. Providers who can afford to serve trillion-parameter models at scale hold a structural advantage that open licensing can't erase.

K3 also didn't arrive alone. It lands in what may be the most concentrated wave of Chinese frontier model releases yet. Alibaba previewed Qwen3.8 Max, which the tech giant said has 2.4 trillion parameters and ranked second only to Anthropic's Fable 5 1. Zhipu, also known as Z.ai, rolled out its GLM-5.2 model last month with rapid adoption worldwide 1.

That density has investors on edge. K3's arrival pressured stocks of US technology companies on concerns that affordable Chinese models could undercut pricing and demand at American AI firms, even as US restrictions already block China from some of the most advanced chips 1.

The takeaway for builders and investors is concrete. Open models are matching or beating closed systems on specific benchmarks, but the gap between a model's capability and the infrastructure to deliver it is widening. Moonshot's 48-hour freeze is evidence that open-weights success creates its own choke point. The companies that solve inference and serving infrastructure for trillion-parameter models, not just the ones that train them, are where the next opportunity sits.

Kimi K3 proved the demand is real. It also proved the pipes aren't wide enough yet.