The extended downtime is unusual for DeepSeek, which has kept its services up without many glitches for over a year.
China’s AI darling DeepSeek suffered one of its longest outages since the viral rise of its R1 and V3 models early last year.
The start-up’s status page noted an initial incident at around 9:30 pm local time (2:30 pm GMT+1), which took nearly two hours to resolve. Subsequently, in the early hours of Monday (30 March), a different incident was flagged, which took until 10:33 am local time (3:30 am GMT+1) to be fully resolved.
DeepSeek said the incident took 7 hours and 13 minutes to be resolved. This is one of the longest outages to affect the company’s web chat services. The cause of the outage remains unclear and DeepSeek has not responded to queries put forth by the press.
The extended downtime is unusual for DeepSeek, which has kept its services up without many glitches for over a year. In the initial days after launching R1, DeepSeek’s API services suffered a few outages, a number of which lasted a whole day. Overall though, the platform has kept its services running 100pc of the time for several months in a row.
A year after taking the AI industry by storm with its large language model, DeepSeek is readying itself for a major update with V4. According to reports, the new V4 is a multimodal model with picture, video and text-generation features. V4 was reportedly trained with the help of Chinese AI chipmakers Huawei and Cambricon.
Rival Chinese AI companies have launched a number of new models in recent months in anticipation of DeepSeek, whose new model could outperform ChatGPT and Claude, particularly on tasks that involve long coding prompts. V4 is expected to the company’s most important product release since V3 and R1.
In February, Alibaba unveiled its latest AI model, Qwen3.5. While ByteDance launched its popular Seedance 2.0, Zhipu unveiled GLM-5, MiniMax launched M2.5 and Moonshot released Kimi K2.5.
Last September, DeepSeek launched V3.2-Exp, an experimental model meant to be an “intermediate step” toward its next-generation architecture. The V3.2-Exp model’s performance is comparable to the previous V3.1 model, performing marginally worse on some benchmarks.
Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.
