Users say DeepSeek is limiting app and web use over a computing shortage
DeepSeek is now capping daily use and input length on its free app and web version and closing long-unused conversations, users say. DeepSeek has not confirmed the changes.
Since 15 November, users of the Chinese AI company DeepSeek have been sharing notices in which the service blames the limits on insufficient computing capacity, Pandaily reported. Chinese outlets including STAR Market Daily and IT Home have covered the story, and Pandaily and Lianhe Zaobao, among others, have summarised their reports.
Long conversations can no longer be continued if they have not been used for more than seven days, the reports say. Some users have reported a 72-hour limit for unused conversations, Lianhe Zaobao wrote. A single input is limited to about 40,000 Chinese characters. When the daily quota runs out, the service says the usage limit has been reached because of server congestion.
The figures are based on users’ observations. DeepSeek has not published any rules or said whether the limits are permanent. The company has no paid version for consumers.
Separate limits for the API
The restrictions apply to the app and web version. The API used by developers has its own rate limits and usage-based billing, according to a report compiled by Guangming Daily. Chinese outlet Chao News advises users who work with long texts to split documents, shorten conversations or switch to the API.
Unused long conversations are closed to free up KV cache, according to an industry analysis cited by Jiemian News. This cache, held in GPU memory, stores intermediate data the language model has computed from earlier text in the conversation, so the model does not have to recalculate it for every response. A long conversation makes the cache grow and, even when idle, takes scarce memory away from active users.
Demand grew and disruptions mounted
On the weekly ranking of OpenRouter, which brings together APIs for many models, DeepSeek’s Flash model processed 33.6 trillion tokens between 19 and 25 October, 801 per cent more than the previous week. Jiemian said the figure refers to V4.1-Flash-1017, the version updated on 17 October. The older V4.1 Flash ranked second the same week with 18 trillion tokens.
DeepSeek’s status page has logged eight disruptions since October, Pandaily, Chao News and Jiemian reported; Guangming Daily counted at least five. Most were slowdowns rather than full outages, Jiemian said. The longest disruption occurred on 15 October and lasted 4 hours 52 minutes, Chao News reported.
On 16 November the status page showed all services operating normally. Qianjiang Evening News, cited by Pandaily, concluded from this that the limits are capacity management rather than a response to an outage. Financial outlet investingLive sees them as a rare visible sign of how scarce advanced computing power may be for Chinese AI companies under US export controls.
DeepSeek has managed demand before. In February 2025 it suspended API balance top-ups, and since August 2026 the API has cost half the peak-time price during off-peak hours. In late May the company restricted regenerating and editing responses, investingLive reported.
The day before the reports, developers found quota management and references to the name Coding Plan in the code of DeepSeek’s website, Pandaily reported. The company has not announced a paid service, and the code has not been linked to the limits users have reported.