A rush of demand has pushed Moonshot AI’s computing resources close to capacity, prompting the Chinese startup to suspend new subscriptions for its Kimi K3 model. Launched on July 17, 2026, K3 is described as an open-weight system with 2.8 trillion total parameters. The company reportedly paused new consumer sign-ups two days later while continuing to prioritize existing paid users. The suspension is a capacity-management measure following the launch-driven rise in usage, not a product withdrawal.
K3’s headline parameter count does not mean all 2.8 trillion parameters are used for every response. Technical documentation describes a mixture-of-experts architecture with about 104 billion parameters active per token, using 16 active experts and two shared experts from a pool of 896. The model is also reported to support a one-million-token context window and visual input through MoonViT-V2. Moonshot released downloadable weights on Hugging Face around July 27-28, with technical documentation recording more than 1.1 million downloads.
The sign-up pause followed a sharp increase in activity after launch. Moonshot said requests over the preceding two days had far exceeded forecasts, bringing its GPU clusters close to capacity. The company has prioritized current paid users to maintain service and plans to offer two membership options, including a coding-focused tier, to match different workloads with available computing resources. Coding and agent-style tasks can require repeated model calls, raising the computing power needed to serve users at scale.
Moonshot is restructuring its offshore holdings in preparation for a potential Hong Kong initial public offering. Reuters has reported that the company is working with financial advisers, including Goldman Sachs and China International Capital Corp. Moonshot and Goldman Sachs declined to comment, while CICC did not respond to inquiries. The reported restructuring involves dismantling a Cayman Islands holding structure and redomiciling the company in mainland China, a process linked in coverage to Chinese requirements for overseas listings. Market reports have described a possible Hong Kong filing in the first quarter of 2027, following a further private financing at an approximately $50 billion valuation. Those figures are reported plans, not confirmed IPO terms.
Founded in 2023 by Yang Zhilin, an AI researcher with academic roots at Carnegie Mellon University, Moonshot has become a major focus of investment in China’s AI sector. The company raised about $1 billion in February 2024 at a reported $2.5 billion valuation, with Alibaba described as its largest investor. Its valuation was reported at approximately $4.3 billion by December 2025. In May 2026, Moonshot secured more than $2 billion, with Meituan among its backers; China Mobile and CPE also participated. That brought total fundraising above $5.5 billion. Coverage put the company’s valuation near $20 billion after the financing; it had reached $30 billion by June, information from figures reported at the time. Separate reporting described a $3.5 billion July financing at a $35 billion valuation, with Tencent among the investors; IDG Capital and China Mobile also participated.
The rapid uptake of advanced models such as K3 is intensifying competition among Chinese technology companies seeking to close the gap with U.S. leaders. Moonshot’s rivals, including DeepSeek, have also turned to external investors to expand their infrastructure. K3’s reported architecture helps explain the operational challenge: although its mixture-of-experts design activates only a portion of its total parameters for each token, delivering a large model to many users still requires large computing capacity.
Kimi K3 is available as an open-weight model, allowing users to download and customize it, although the hardware requirements make independent hosting challenging for many organizations. Moonshot says K3 matches or exceeds leading U.S. models on certain technical benchmarks, a claim echoed by independent evaluations. The launch follows the company’s multimodal Kimi K2.5, released in January 2026 with native vision capabilities through a MoonViT encoder described as having roughly 400 million parameters. K3 extends that product line with a substantially larger mixture-of-experts system and a longer context window, positioning it for coding and reasoning tasks, including extended knowledge work.
Competition is also reflected in Alibaba’s debut of Qwen3.8-Max-Preview, a 2.4-trillion-parameter model, on its AI platforms. U.S. export controls on advanced Nvidia chips continue to constrain Chinese access to high-performance computing hardware. For Moonshot and its competitors, the ability to secure investment and computing resources will shape how quickly they can scale model access while managing the costs of serving increasingly demanding workloads.
China’s AI industry has grown rapidly, with startups such as Moonshot at the center of technological competition and investment. K3’s brief path from launch to a pause on new consumer sign-ups illustrates the infrastructure pressure that can follow sudden demand for a high-profile model. The company’s fundraising and reported Hong Kong listing preparations place that capacity challenge within a wider effort to finance expansion and compete in the global market for advanced AI systems.
