Economic Observer reporter Ren Xiaoning

On the early morning of October 26th, Kimi, a subsidiary of Moon Dark Side, released an apology and compensation to users regarding recent user experience issues. This is the third time Kimi has apologized to users since April this year.

The reason for this apology is that users encountered "traffic restrictions and unnecessary quota consumption" when using the new model K3.1 Pro. The compensation plan is to reset all subscription members for each level, and members within the validity period will receive 25% of the monthly total amount, equivalent to one week's usage.

Kimi's quota refers to the fixed amount of tokens that users can use each month after paying. According to the updated domestic pricing on Kimi's official website on October 24th, the monthly subscription for members is 49 yuan, 99 yuan, 199 yuan, 699 yuan, and 1399 yuan respectively. These member functions share the same monthly quota pool and are deducted based on actual token consumption. Exhausting a certain function will affect other functions, and the unused quota in the current period will not be carried over to the next month.

Chen Jing, the R&D Director of Asia Visual Technology Co., Ltd., is a loyal user of Kimi and spends 199 yuan per month to subscribe to Kimi. On October 15th, he spent 1908 yuan to renew his membership for one year. The next day, he asked K3 to generate a 5-minute video and ran the task for 20 minutes. The credit was reduced from 100% to 0%, but the video was not produced.

I never used to worry about my credit limit running out before, "he said. Now when I open Kimi, I feel an inexplicable sense of panic, afraid that my credit limit will be depleted too quickly.

A person close to Kimi told the Economic Observer that the reason for this incident was due to some configuration issues with the new model, which tends to assign work to sub agents to complete tasks and bring more output. Now Kimi has added the option for users to independently choose whether to use sub agents.

Kimi's official apology letter acknowledges that 'the problem was with us and we should have fixed it earlier', and states that the configuration has been adjusted to reduce unnecessary sub agent calls and output lengths, allowing the same amount of resources to complete more tasks.

Chen Jing has received a compensation amount of 25%. But his sense of panic did not disappear, 'I dare not try big tasks again'. He did some simple chat tasks with K3.1 Pro and found that the consumption rate of the quota was still faster than before the model upgrade.

The credit limit is running out quickly

On September 27th, Chen Jing encountered the problem of Kimi's credit limit being quickly depleted. On that day, he used a longer script to ask Kimi to make an 8-minute video, and Kimi immediately used up her monthly quota. At that time, he didn't realize that Kimi might have a problem and thought it was his own carelessness.

Until October 16th, when he encountered the same problem again, he felt that it was probably not his fault. At that time, he had just renewed his annual fee and the quota was almost full, but Kimi still hadn't made any videos.

Chen Jing has been using Kimi for two and a half years, and he fully recognizes Kimi's modeling and product iteration capabilities. He can also clearly feel the improvement in Kimi's chat quality and agent task quality.

But now he has the idea of whether to switch to a different model. After running out of credit, Kimi tried Tencent's WorkBuddy+DeepSeek-V4.1-flash and only relied on the points given by the platform to create 8 videos of 5 minutes or more.

After discovering Kimi's problem for the second time, Chen Jing went to customer service to complain, but the customer service did not restore his credit limit at that time. The customer service explained that more than 20 intermediate result videos were generated during the task process, and there were corresponding expenses.

Technically, it can be understood, "Chen Jing said. Kimi generates videos using large models instead of searching for materials online and splicing them together using CPUs, which naturally incurs much higher costs. But from a product perspective, he cannot accept this experience. If he spends his money and doesn't receive the video, the monthly membership fee will be invalidated.

He hopes that while Kimi continues to challenge the high performance of the model, it should also focus on optimizing the user experience, reducing token consumption, or at least giving users a prompt when they are about to consume a large amount of tokens.

Chen Jing compared other Agent products: When encountering heavy tasks, WorkBuddy will inform users in advance how many points will be consumed for this step, ask if they accept it, and if they do not accept it, they will not do it. On Kimi's side, 'the task is to run silently, and as soon as I see something is wrong and want to terminate it, it's already too late.'.

Chen Jing is not the only user who has discovered the issue with Kimi's credit limit. After the launch of K3.1 Pro on October 24th, similar feedback erupted on social media platforms.

A user who purchased a 199 yuan package said that with just one sentence of "review project architecture," Kimi pulled up 5 sub agents, and the limit was exhausted after more than ten minutes; Someone left a message under Kimi's apology post saying that on the morning of the 26th, a simple question burned 25% of the credit limit; Another member who purchased the 199 yuan package said that they used to only use half of their quota in a week, but now they can reach the limit with just one use.

The bottleneck currently restricting Chen Jing's replacement of large model products is the cost of migration and usage habits. He has accumulated two and a half years of chat records and content in Kimi, and many complex tasks are done in Kimi chat boxes. Changing them all at once is equivalent to starting over.

Why is there excessive consumption?

According to the data from the Artificial Analysis model page of a third-party evaluation agency, K3.1 Pro has accumulated approximately 150 million tokens in the Intelligence Index test, while the median of the tested models is 87 million, which is about 1.7 times the median level. Artificial Analysis is therefore labeled as "very verbose" on the model page.

After entering the subscription system, this redundancy is further amplified. Subscription users purchase a fixed pool and are particularly sensitive to how long a task will take and how much money it will cost. The better the model is at multi-step process search, calling tools, writing code, and self checking, the longer the task time, and the more likely it is to hit the "quota wall".

Zhang Faen, a senior artificial intelligence engineer in Beijing, believes that there are two main reasons for Kimi's problems: firstly, the company lacks computing power, and secondly, the model is too large.

At present, Kimi has the largest parameter among domestic models, reaching the 3 trillion level, and the cost of a single token is naturally more expensive. For the question of why users feel that their quota consumption has increased after the model update, he believes that the most likely reason is that post training has made the model think longer, "doing the same thing costs more tokens".

In addition, Kimi's computing power is more scarce compared to its peers. According to his understanding, DeepSeek and Zhipu have already completed inference using domestic cards, while Kimi is currently mainly using overseas chips with relatively limited computing resources. The combination of two factors can easily lead to problems when the new model is launched.

However, he believes that the problems Kimi is currently facing will not last long. He estimated that after K3.1 Pro, there will be smaller parameter versions, just like DeepSeek has both Pro and Flash versions. After the Flash version is released, the token cost will decrease significantly, possibly up to 10 times, and the work will not be bad. ”

Prior to this, he suggested that Kimi users could switch to a low thinking mode to save tokens, at the cost of a decrease in work quality.

The most eye-catching model, apologizing the most

Among domestic large model companies, the dark side of the moon ranks at the top, and there are reports that its latest valuation has reached $50 billion. Since the beginning of this year, Kimi has embarked on a fairly dense model iteration curve, and after the release of the K3 model, it has received a lot of attention from the industry.

Goldman Sachs stated in a research report that K3 marks a key milestone for the global popularization of China's open-source model intelligence, and Chinese AI companies are transitioning from "cost efficiency" competition to "pricing power". Morgan Stanley also commented that the release of K3 indicates that China's big model is catching up with American leaders in terms of model size, performance, and pricing in all aspects.

The latest release of K3.1 Pro model is another significant improvement in quality after nearly two months of silence for domestic models. On the Intelligence Index of Artificial Analysis, K3.1 Pro ranks fourth globally and first in open source, jumping 12 places from K3.

However, every time the model's capabilities are recognized, Kimi products are also questioned by users.

From April to October, Kimi apologized three times. The three apologies all occurred after the new model was launched, and the reasons for the apologies were related to issues such as traffic restrictions and quotas.

Kimi's first apology was on April 22nd. On April 20th, Kimi released and open-source a new model K2.6, which focuses on long-term coding. Kimi apologized afterwards, stating that due to the sudden surge in traffic, some users encountered situations where members queued up and functions were temporarily unavailable; At the same time, there was a deviation in the backend system's calculation of Agent quotas, resulting in some users' rights being mistakenly deducted. At 20:30 that evening, Kimi restored all users' monthly quotas to 100% and cleared their usage to zero.

The second apology was on July 19th. Kimi launched the Kimi K3 on July 16th, which was the world's largest open-source model at the time. Within 48 hours of publication, the number of user requests significantly exceeded the estimate. Kimi released a statement titled "To Kimi Users: Explanation on Computing Shortage and Membership Suspension" overnight, announcing the suspension of new user subscriptions on the C-end, leaving all existing computing power to existing members, and apologizing solemnly to friends who were "full of expectations but did not receive the expected experience".

The latest occurrence occurred in October. On the early morning of October 24th, K3.1 Pro was released with a context window of 3 trillion level parameters and 1 million tokens, mainly enhancing programming, 3D creation, financial research, and long-range agent capabilities.

A person who has served as a senior executive in a top domestic model company told reporters that Kimi is a respected company in the field of large models, both in terms of technical strength and investment. From the current perspective, there is a certain gap between Kimi's aggressive pursuit of technological breakthroughs and user demands. However, he also understands Kimi because the current big model company is in the stage of 'the more users, the more losses'. He also believes that Kimi's next step should be to launch small parameter model products to achieve a balance between commercial purposes and user needs.

(Author Ren Xiaoning)