Moonshot AI has temporarily paused new Kimi subscriptions after demand for its Kimi K3 model pushed the company’s computing capacity close to its limit within 48 hours. Existing members remain supported, while new subscription places will reopen gradually as more capacity becomes available.
The July 20, 2026 update follows Kimi K3’s launch on July 16. The Beijing-based company confirmed that its new flagship model has 2.8 trillion parameters, native visual understanding and a one-million-token context window.
Why did Kimi pause new subscriptions?
Moonshot said demand was much higher than expected and brought its computing clusters close to maximum capacity. The pause applies to new consumer subscriptions rather than the entire Kimi service.
The company is prioritising existing subscribers and adding computing resources. It is also separating Kimi Membership for web, app and general work from Kimi Code Membership for programming, allowing capacity to be divided between different workloads.
What can Kimi K3 do?
Kimi K3 is designed for long-running coding, research, document analysis, visual tasks and complex reasoning. According to the official Kimi K3 documentation, it uses Kimi Delta Attention, Attention Residuals and a mixture-of-experts architecture that activates 16 of 896 experts for each token.
The one-million-token context window allows the model to process large codebases and lengthy documents. A large context limit does not guarantee accuracy, however, and important outputs still require human verification.
Kimi K3 price and availability
The Kimi K3 API costs $3 per million input tokens and $15 per million output tokens. Cached input is priced at $0.30 per million tokens, excluding any applicable taxes.
The model is available through Kimi’s hosted products and API. Consumer membership access is separate from API billing and is currently limited for new subscribers.
When will Kimi K3 become open weight?
Moonshot plans to publish Kimi K3’s full model weights by July 27, 2026. This would allow developers to inspect, customise and deploy the model on compatible infrastructure, subject to its final licence.
Running Kimi K3 locally will not be practical on a typical personal computer. A model with 2.8 trillion parameters requires specialised hardware, substantial storage and significant operating resources.
Kimi K3 versus Claude Opus 4.8
Kimi K3 is being compared with Anthropic’s Claude Opus 4.8 because both target coding, reasoning and long-running professional work. Early evaluations show Kimi performing strongly in some frontend and agentic coding tests, but results can change according to the prompts, tools and testing methods used.
The overview of Claude Opus 4.8 and its autonomous-agent capabilities provides useful context on Anthropic’s approach. Claude is a proprietary hosted system, while Kimi K3 is moving toward downloadable weights.
Parameter count alone cannot identify the better model. Accuracy, reliability, speed, tool performance, safety and total operating cost are more meaningful comparisons.
Why the Kimi K3 launch matters
Kimi K3 arrived as Alibaba introduced Qwen3.8 Max, adding to China’s competition with leading US AI developers. The wider US-China dispute over AI development and model distillation shows that this competition now includes intellectual property, chip access and national security concerns.
The new models do not prove that one country or company has secured an overall AI lead. Company benchmark claims still require independent verification across factual accuracy, multilingual performance, cybersecurity and real-world professional tasks.
Moonshot’s next important milestone is the planned July 27 weight release. Until then, the subscription pause highlights the immediate challenge of serving a powerful model to a rapidly growing user base without reducing reliability.
Make Swikblog your go-to source on Google for reliable updates, smart insights, and daily trends.
