DeepSeek suffered an epic 13-hour outage and trended on the hot search overnight. Is the web version updated and V4 really here?

[Introduction] 13 hours of the darkest shutdown, an overnight "disconnection" triggered wails across the entire network... The longest outage in DeepSeek's history shocked the world. All signs indicate that V4 may have quietly made its debut. The leaked benchmark tests suggest that perhaps this is the eve of the apotheosis of the domestic large - model star!

Just yesterday, the DeepSeek outage incident has become a hot topic overseas.

From 21:35 on the evening of March 29th until 10:33 on the morning of the 30th, a full 12 hours and 58 minutes. DeepSeek, the star of domestic large models, experienced its longest server shutdown since it became popular in early 2025.

The topic on Weibo went viral, with the popularity directly reaching over 950,000.

The overseas developer community went crazy, and foreign media specially published articles to report on this outage incident.

Amid the wails across the network, the anticipation for DeepSeek's new model has never been so strong!

What on earth happened to DeepSeek? The new model has not been released for a long time, and every move of it, even the slightest clue, has attracted the attention of netizens around the world.

This discussion post about DeepSeek's new developments has gone viral on Reddit.

It has done nothing for several months, yet it can still tug at the hearts of all netizens. Perhaps only DeepSeek has such a status.

However, is this really just a simple outage?

All the clues indicate that this may not be an accident, but a grand prelude. DeepSeek V4 may have quietly sneaked into our computers during this darkest moment!

DeepSeek, please come back soon

Thirteen hours made many people realize that they have become inseparable from it.

How serious was this outage? Foreign media used a term in their reports: Major Outage.

Looking back at DeepSeek's rise to fame, whether it was the R1 version that shocked the world at the beginning of the year or the V3 version that swept the benchmark tests, although DeepSeek would occasionally experience glitches, the web - based service had never been interrupted for more than 2 hours.

This time, throughout the whole night, both the web and App versions turned gray.

The red warning on the official status page flashed all night, and the engineers worked overnight to repair it.

Moreover, it's not just Chinese users; the overseas developer community was also in an uproar.

In January this year, the DeepSeek API also crashed, but such a long - term shutdown of the web version, which serves as the front - face, is extremely rare.

This "disconnection" made many people suddenly realize that AI is no longer an optional toy but a necessity for productivity.

When that dialog box disappeared, how many office workers' workflows were directly stuck? It felt like losing half of their brains.

Is V4 quietly undergoing a gray - scale test?

While everyone was complaining about the outage, the first batch of users who logged into the system found that something big was happening!

If you successfully used the restored DeepSeek this morning, please carefully observe its output logic.

On social media, a large number of developers have excitedly shown evidence that DeepSeek V4 may really be here.

First of all, its coding style has changed overnight.

Many senior developers found in the Zero - shot coding test that DeepSeek's current output structure is clearer than a few days ago, and the logical rigor has made a qualitative leap.

Although in the dialog box, it still modestly calls itself "V3", but that sensitivity of the underlying logic is probably not something that fine - tuning can bring.

For drawing a pelican riding a bicycle in SVG, it looked like the one on the right a week ago, but yesterday it had evolved into the one on the left.

Also, pay attention to the self - introduction circled in the red frame. A week ago, the self - introduction only said that it was a pure - text AI assistant, but now in the self - introduction, it specifically emphasizes the version number and mentions that it is DeepSeek V3.

Another netizen found that DeepSeek has been upgraded in the web application, and the output effect has been greatly improved.

This experience really seems to match many people's first impression of V4.

Moreover, the UI has also quietly changed.

Careful users found that the upload button interface on the App version has been updated. And such small changes usually indicate a huge adjustment of the back - end architecture.

Another piece of evidence for DeepSeek's update is that it knows who the current US president is, so the cut - off date on the DeepSeek web version is 2026.

However, why doesn't this new model change its name?

Many people guess that the official is probably conducting a gray - scale test.

It's very likely that some capabilities of V4 (or a version distilled from V4) have replaced the existing V3 interface, but the label hasn't been changed yet. In the terminology of the AI circle, this is called "invisible evolution".

V4 benchmark test leaked

If the style change is just a guess, then the following one is quite astonishing.

Recently, a set of "DeepSeek V4 benchmark data" has begun to circulate in the circle. Although it hasn't been officially confirmed, it contains a lot of information -

According to the information from the technical community, the specifications of V4 can be described as an aesthetic of brute force.

First of all, the scale of the V4 Lite version reaches 200B parameters.

In addition, it also supports a 1M Token context window. It can instantly understand the entire underlying source code of a project.

The multi - modal package of V4 covers text, images, and videos.

The most revolutionary one is the update of the mHC architecture. Through this new architecture, V4 can be scaled up to the level of 1 trillion (1T) parameters!

On the HumanEval list, V4 reaches as high as 90%, and on the SWE - bench, it exceeds 80%. Moreover, the code performance of V4 is higher than that of V3.2 and all known competitors in the current market.

According to the online information, the following is a performance comparison chart between V4, Claude 3.5 Opus, and GPT - 5.3.