Kimi Apologizes Three Times in Six Months: The Stronger the Model, the Harder the Business?

At the end of October, OpenAI once again adjusted its quota rules, responding to users' complaints that the quota for the Pro 200 plan had been cut. After a sentence about "receiving everyone's kind feedback," Codex lead Tibo announced that usage limits for all Pro and Business Premium plans would be doubled. In the community, cheers kept coming over this.

Something similar happened to Kimi recently, but the picture seems even more "brutal."

Less than 12 hours after Moonshot AI's new model Kimi K3.1 Pro went live, controversy followed one after another. On one side, an overseas blogger posted an exposé that the promotional templates it provided also required hiding paid partnership labels during promotion; on the other, users concentrated their complaints on rapid quota consumption: not long after a task started running, they found they were already close to the usage limit.

Two days later, Moonshot AI publicly apologized for K3.1 Pro's rate limiting and unnecessary quota consumption, restored 25% of the monthly quota in one go for all plan subscribers, and said it would reduce unnecessary sub-agent calls and output length.

Although this is Kimi's third apology within half a year, this "quota incident" has not truly been put behind it.

The "quota black hole" continues

Among domestic large models, Kimi has always been a special presence. Relying on long-text processing and text creation capabilities, it has not only built up a group of loyal users but is also capable of going head-to-head in global frontier model competition with overseas giants that have massive computing power and capital reserves.

But since the beginning of this year, complaints that its subscription plans run out too quickly have never stopped. "Those who can't afford it can't use it, and those who can afford it find overseas models more cost-effective." A heavy AI user said that at the same price, the quota Kimi provides is far less than that of GPT and Claude. Another user compared it with other models: "199 yuan topped up on DeepSeek is enough for a month, but paid to Kimi it only lasts a day."

For these users, large models are not a chat tool they open occasionally, but a productivity product embedded in their daily workflows. Therefore, their perception of quota consumption is very direct.

After K3.1 Pro launched, this contradiction became even more prominent. The new model saw a clear round of upgrades in capabilities such as coding, long-horizon tasks and 3D modeling, and the market response was quite positive. Artificial Analysis gave it an intelligence index score of 55, ranking fourth globally and making it the highest-scoring open-weight model on the list. The third-party evaluation agency ModelCompare even said because of this model's appearance that "the AI gap between China and the United States has narrowed from several months to just 31 days."

Looking only at API pricing, Kimi is still working hard to maintain the cost advantage domestic models have always had. K3.1 Pro was cut by about one-third compared with the previous generation, with input dropping from 20 yuan to 13 yuan per million tokens and output from 100 yuan to 68 yuan. But for users who buy subscription plans, even though model capabilities have improved, in actual use they need to calculate their quota more carefully.

A long-term Kimi subscriber found after comparison that although the unit price is lower, the new model consumes more tokens to complete a single task, so the overall cost works out about the same. According to Artificial Analysis testing, K3.1 Pro accumulated about 150 million tokens of output to complete the full Intelligence Index, while the median among the models tested was only 87 million tokens. Converted, K3.1 Pro's token output is about 72% higher than the median level.

Actual task cases are even more startling. In the community, a user who subscribed to Kimi's 199 yuan plan only sent the instruction "review the project architecture," and K3.1 Pro spun up 5 sub-agents; more than ten minutes later, the 5-hour quota window was already used up. Faced with the classic "pelican riding a bicycle" test, K3.1 Pro generated for 1 hour and 47 minutes, consuming 87% of the 5-hour quota.

What is more troublesome is that the quota compensation also failed to fully calm users' discomfort. Over the past few days, users have continuously reported that even with quota, they cannot use Kimi now.

"Kimi is in peak hours, please try again later or upgrade your membership for faster access." Even after paying for a subscription, this maddening prompt still appeared many times in the chat boxes of a large number of users. Among them were not a few users who had already upgraded their plans once because their quota was insufficient.

Kimi has pushed the model to a higher capability ceiling, but what it brings users is no longer just the quota consumption problem, but also the gap between peak-hour service capacity and paid rights.

"Afraid to give enough"—has Kimi started touching memory?

To ease quota pressure, Kimi seems to have begun adjusting the memory feature it was once most proud of. On August 10, without notifying users in advance, Kimi updated its memory architecture and introduced a "dream memory" mechanism that consumes paid quota, automatically reading conversation content in the background and writing it into the memory bank.

Now, multiple users have reported that Kimi's memory space cannot load properly these days, suspected to be because the company is upgrading and modifying the memory mechanism. The current manifestation is that the new version of memory tends to uniformly summarize past memory instructions and distill them into user preferences, rather than executing memory content item by item as before. The benefit is that token consumption may be lower. But the cost is that the originally more detailed memory capability will also be affected.

But ultimately, this can only alleviate part of the consumption. The cost problem Kimi faces first comes from the model itself.

K3.1 Pro is a 3-trillion-level parameter model. Previously circulated industry information about the K3 series architecture showed that its model scale is about 2.8T parameters, with about 104B activated parameters. By comparison, some models with relatively smaller parameter scales and higher activation efficiency have more advantages in engineering cost control. Vendors such as DeepSeek and Zhipu are all seeking a more suitable balance between capability and cost through model architecture and inference engineering optimization.

Second is the inference consumption brought by complex tasks. Kimi K3.1 Pro enables thinking mode throughout by default, outputting longer chains of thought and more intermediate reasoning content. Long context increases the computational overhead of input processing, and complex reasoning and multi-turn Agent calls will further push up total consumption. If an Agent task requires multiple model calls, running multiple sub-Agents, and then repeated checking and revision, the resources ultimately consumed may far exceed the single question as intuitively perceived by the user.

For Kimi, what really needs to be proven may not be whether it can continue to make model parameters larger, but its ability to raise inference efficiency while upgrading capabilities. A developer who follows Kimi said that rather than blindly participating in competition among high-end models, Kimi should perhaps invest resources in developing lighter and more efficient models.

This is not a problem for Kimi alone.

All vendors that provide frontier model services at a fixed monthly fee face the same structural tension: improvements in model capability are often accompanied by synchronous or even superlinear growth in inference-side overhead, while subscription prices are rigid and user expectations are that "if I pay, it should be enough." The difference is that some companies have made token efficiency part of their product capability, some have made capability breakthroughs into brand assets, and subscription users are caught in the middle, paying runtime costs beyond expectations for "powerful intelligence."

In addition, Moonshot AI's computing resources rely heavily on external supply. Bloomberg once reported that it has a computing agreement with Alibaba and can use a cluster composed of about 20,000 Nvidia chips, with related computing power accounting for a considerable portion of Moonshot AI's overall computing resources. Previously, within 48 hours after K3 was released, Kimi suspended new subscriptions because GPU capacity was approaching its upper limit.

When a company's computing power itself is in a tight state, opening generous quotas to consumer subscription users is financially irrational.

Domestic models, whose money should they earn?

If a company's computing supply is limited, it must decide a practical question: where should resources be prioritized? Training and iterating the next-generation model, investing in real-time inference services for a large number of individual users, guaranteeing developers' API call needs, or providing subscription users with more ample quotas?

Behind different choices correspond different commercialization priorities. Kimi's quota controversy also exposes Moonshot AI's true positioning of consumer subscriptions under computing constraints.

From August to September 2026, multiple media outlets reported that Moonshot AI was negotiating with Microsoft, Amazon, and Google, planning to list K3 on Azure, AWS, and Google Cloud, and to take a share of up to 30% of the revenue generated from K3 services sold through these platforms. In addition, K3's open-source license is not the standard MIT license but terms drafted by Moonshot AI itself: MaaS hosting providers with annual revenue exceeding $20 million must separately sign a commercial licensing agreement.

Putting these two things together, it can be seen that Moonshot AI is exploring a commercial path different from simply operating consumer subscriptions: compared with building large-scale consumer inference infrastructure itself, "wholesaling" model capabilities on a priority basis to cloud vendors with computing power and distribution channels is far more in line with its resource situation than "retailing" to millions of subscription users.

And when consumer subscriptions play more the role of user entry point, product feedback, and model capability display, rather than the main source of revenue, then their quota pool may differ from user expectations.

At present, domestic large-model companies are also looking for their own more suitable ways to make money.

DeepSeek's competitiveness comes largely from the combination of model capability, inference efficiency, and an open ecosystem. At present, almost all of its revenue comes from model API call charges, and the chatbot App remains free and ad-free. According to The Information, DeepSeek's annualized revenue run rate has exceeded $1 billion, more than doubling from less than $500 million a few months earlier. CEO Liang Wenfeng clearly stated at a recent investor meeting that the API price increase has not caused customer churn, and user demand remains strong.

Zhipu, meanwhile, is undergoing a business model transformation, shifting from reliance on localized deployment to API subscription services. Its financial report for the first half of 2026 shows that revenue from the Zhipu open platform and API business accounted for 86.5%, a year-on-year increase of 2735.7%, while the share of localized deployment revenue fell from 84.8% to 13.5%. Behind this change, the explosive demand for Coding is an important driving force.

MiniMax, which started with overseas ToC products, is trying to broaden the boundaries of commercialization. In addition to continuing to build out models, APIs, and products for individual users, it is also increasing efforts to expand the B-end market, extending its capabilities to video generation as well as Agent, Reasoning, Coding, and other directions. Its approach is to bring model capabilities into more enterprise and developer scenarios and avoid excessive reliance of revenue on a single product or customer group.

Compared with these model companies' direct monetization paths, internet giants have another advantage: they already have mature product matrices, user scale, and enterprise customer resources, and can embed large models into existing ecosystems and then seek revenue through product value-added and enterprise services. For example, Tencent is trying to use large models to transform its products and office scenarios. Products such as WeChat and WeCom have successively integrated model capabilities such as Yuanbao, Hunyuan, and DeepSeek, and AI office agents such as WorkBuddy are also accelerating user expansion.

After integrating product capabilities such as QoderWork, MuleRun, and Wukong, Alibaba launched the one-stop enterprise-level general Agent platform Qianwen Office. At the Yunqi Conference in September, DingTalk completed a commercial version upgrade. On the basis of retaining the free standard edition, it added four paid versions: AI Basic Edition, AI Advanced Edition, AI Premium Edition, and AI Flagship Edition, and fully integrated Qianwen Office capabilities.

Low prices and traffic alone cannot sustain a long-term business. Model capabilities can continue to break through upward, but users' willingness to pay will not grow indefinitely. The real question is whether each company can find a business model that matches its own technical capabilities, computing resources, and market position.

Far more than just Kimi are standing on the edge of a cost cliff.