DeepSeek Adjusts App and Web Version Quotas, Users Give One-Star Negative Reviews
DeepSeek recently adjusted usage quotas for its App and web version, no longer supporting conversations with no interaction for more than 7 days, lowering the single-input text limit to 40,000 characters, and imposing a daily conversation quota; once the quota is exhausted, users must wait until the early hours of the next day for it to reset. The move has sparked user dissatisfaction: its rating on the vivo app store has dropped to 3.1, and there has also been a large amount of complaining in the comment section of its Xiaohongshu account. DeepSeek has not yet responded. The article states that the quotas are related to computing power pressure; according to OpenRouter statistics, from November 9 to 15, DeepSeek V4.1 Flash was called 59.2 trillion tokens, ranking first globally.
Why read this
AI score 62
The original article clearly explains the computing power consumption structure behind the quotas, allowing readers to understand the resource trade-offs between the free consumer-facing (C-end) service and the B-end API.
Topics
Body · AI translation
DeepSeek Sets Usage Limits, Freeloaders Panic! Furiously Blast Liang Wenfeng Tech Daily Push November 18, 2026 11:50:05
Author | Zhao Zhishan
Editor | Zhou Weipeng
Recently, DeepSeek quietly adjusted the usage quotas for its App and web version.
It no longer supports conversations with no interaction for more than 7 days, while the single-input text limit has been lowered to 40,000 characters, and daily conversation quotas are now restricted.
Once a user reaches the daily usage limit, they can only wait until the early hours of the next day for it to reset.
Unexpectedly, this move sparked widespread dissatisfaction among users, who stormed into app stores to give DeepSeek one-star reviews and even furiously cursed out Liang Wenfeng.
vivo shows its rating has dropped to 3.1, the lowest among mainstream domestic AI applications.
The comment section of DeepSeek's Xiaohongshu account has also been overrun.
Users complained about why there was no advance notice, asking whether the company simply doesn't care about the fate of C-end users, while on the other hand begging DeepSeek to bring back unlimited quotas.
Currently, DeepSeek's Xiaohongshu account has not posted a quota limit notice, nor has it responded to the matter.
I thought it was just Liang the Saint being stingy, but after taking a look, I realized these people are really playing too wildly.
DeepSeek didn't suddenly become stingy; its computing power really can't keep up.
According to OpenRouter statistics, during the week of November 9 to 15, DeepSeek V4.1 Flash was called 59.2 trillion tokens, ranking first globally.
Since October, the platform has experienced at least 5 service anomalies, and the outage on October 15 lasted as long as 5 hours, the most severe one this year.
To preserve B-end demand, DeepSeek can only cut back on the C-end.
If users were only having simple conversations, computing power consumption would be fine, but the key is that many people use it for other things.
Word games, interactive fiction, role-playing, and even human-AI romance—DeepSeek users were really playing this wildly.
And these ways of playing are extremely computing-power-intensive.
Ordinary people finish using a conversation after looking up information, ask a few questions, and then close it, but these users typically open a dozen or dozens of independent conversation windows, each stuffed with tens of thousands of tokens of worldbuilding, character cards, and past plot, with a single session even lasting thousands of rounds.
As long as users don't close these long conversations, they keep occupying HBM memory, ready to continue at any time.
The most expensive part of a GPU is HBM memory, and this practice is equivalent to occupying it for free without using it.
Moreover, every time the model has a conversation, it has to re-read the historical conversation, and many people repeatedly append new settings without summarizing or archiving, only stacking it longer indefinitely.
Once the model forgets the earlier character settings and produces an answer that doesn't match their imagination, users repeatedly revise their questions to make DeepSeek regenerate answers, and some even made it repeat an answer more than 6,000 times!
Like games, text-based play is also easy to get hooked on. Some users claim they spend at least 15 hours a day playing with DeepSeek, which means that apart from sleeping and eating, their day is basically spent on this.
So in the comment section, some users called for a paid entry point on the App, saying that even a paid package would be acceptable as long as there are no limits.
Why doesn't DeepSeek add a payment page and insists on going with API?
DeepSeek's main line is AGI, and the value of the C-end is to let everyone try the model, plant the seed, and ultimately convert them to using the API.
Therefore, its computing power is prioritized for developers who can pay continuously and stably and consume tokens at scale, rather than individual users.
If the App launched a membership, it would have to carve out a separate priority memory pool for paying members, which amounts to building an entire extra system of cache isolation, priority scheduling, and session keep-alive—no small engineering effort.
And this would conflict with its route of prioritizing B-end developers.
In-app subscription membership is also not as simple as adding a button; both iOS and Android take a cut, while API goes through open platform web top-ups and avoids store commissions.
At the same time, it would also need to support automatic renewal, refunds, invoices, membership rights arbitration, and paid service agreements.
More importantly, once it charges users a monthly fee, it is equivalent to promising stable quotas and retained sessions, and its current computing power supply cannot keep up.
So it chose to open up its capabilities and let heavy users build their own frontends and handle billing themselves.
When it went viral during last year's Spring Festival, they never thought about how to retain users, monetize, or build a super app, and to this day the overall App pages have basically not changed much, with no large-scale marketing like other AI applications.
In July this year, in a leaked DeepSeek closed-door exchange meeting, DeepSeek founder Liang Wenfeng said that they had not put much effort into the C-end, and at one point didn't even want to maintain those users, but the users just wouldn't leave.
For users, though, with no ads, no membership, and no payment pop-ups, and a world-class model used for free for two years, they have already formed a habit, and at this point the quota limits are like a blow to the head.
The good news is that next year, after Huawei Ascend supply ramps up, the computing power shortage may be alleviated.
DeepSeek plans to build a data center in Ulanqab, Inner Mongolia, deploying at least 160,000 Ascend 950DT chips.
But Liang Wenfeng has said that using domestic chips to train the next-generation model is the company's highest priority. That is to say, even after Ascend supply comes up, C-end demand will still be last in line.
So DeepSeek heavy users, while there's still time, might as well start using the API early and enjoy the benefits first.
AI Trends