On November 3, the artificial intelligence platform DeepSeek experienced yet another outage, a situation that disrupted both its conversation features and two of its model interface services for a significant period. It wasn’t until nearly two hours later that the service status page indicated that operations had returned to normal.
According to reports from First Finance, on that afternoon, users were unable to access the chat function, with the interface displaying the message, “Server busy, please try again later.” The phrase “DeepSeek crashed” quickly became a trending topic on Weibo, indicating the level of frustration among users.
The outage specifically affected the conversation service as well as the V4 Pro and V4.1 Flash API services. APIs are essential as they allow developers and businesses to integrate model capabilities into their software applications.
During the repair period, there were brief moments when the conversation and some interface services seemed to recover, but these were followed by further declines in performance.
As reported later that evening by the Beijing News, the time frame of the incident was recorded from 4:58 PM to 6:48 PM Beijing time. By the time the publication went to print, the status page indicated that “everything is operating normally.”
In the aftermath of this incident, online commentary reflected a mix of sarcasm and frustration. One user lamented, “DeepSeek crashes so easily.” Another added, “It’s getting worse; the user experience is declining.” Some users mockingly inquired, “What kind of compensation are they planning this time?” Another remark lightheartedly noted, “At this point, we’re almost used to DeepSeek crashing; we can expect a small crash every month and a major one every two months.”
Statistics from the International Financial News reveal that DeepSeek has recorded a staggering 55 outages since the beginning of this year alone. Notably, on October 15, DeepSeek went down at 8:11 AM and did not announce the restoration of services until 1:04 PM, resulting in an outage lasting nearly five hours. This incident affected ordinary users as well as developers and businesses that depended on the API interfaces.
Following the previous outages, the company issued an apology, offering account bonuses equivalent to the consumption amounts of API users from October 8 to 14. These bonuses were intended to offset the costs of future API usage. The company also stated its commitment to investigating the stability of its services in order to minimize the ripple effects caused by failures in individual components.
Analyzing the October 15 outage, Tian Feng, the director of the Critical Thinking Research Institute, explained to The Paper that during the beta testing of V4.1 Flash, each account could manage a maximum of 20 concurrent requests. The platform has also implemented a pricing strategy based on peak and off-peak usage to regulate demand, which suggests that the availability of computing power is quite limited.
Another executive from a cloud computing service for AI echoed this sentiment, calling attention to the need for improved operational and maintenance engineering practices. They highlighted the importance of establishing robust testing, upgrading, and recovery processes to handle such incidents more effectively.
Tian Feng further noted that the more AI is involved in various tasks, the closer the system gets to total interruption during a failure. When tasks are interrupted midway, modifications to documents and calls to tools may need to be redone, which compounds the challenges of recovery.
This ongoing struggle with service reliability underscores a pressing need for the company to address its operational shortcomings and improve the overall user experience for both individual users and enterprises reliant on its technology.