Confirmed! DeepSeek officially announced last night that the web version and APP have been updated to support 100k token context. Nowadays, the whole Internet is squatting on DeepSeek V4.The rumored DeepSeek V4 is getting closer! After several days of grayscale testing, last night, DeepSeek officially announced that it had updated the web and APP sides——

In the new long text model structure test, up to 1 million token contexts are supported.


However, API players will have to wait a little longer. It is still V3.2 and supports 128k context.

This "toothpaste squeeze"-style surprise release has plunged many people into a carnival of reminders. Today, the entire network is waiting with bated breath for the official arrival of V4.



DeepSeek suddenly went cold after update

In the past few days, many people have discovered that DeepSeek's personality has suddenly changed.

The once considerate AI responded with an unusually cold attitude, and some netizens even complained about its "yin and yang" speech.

For a while, "DeepSeek was accused of becoming cold" went straight to Weibo's hot search, and Xiaohongshu and Zhihu were full of people's confusion and complaints.


This controversy originated from a "grayscale update" on February 11, which is the content of this official update.

At that time, netizens discovered that when DeepSeek was asked to make the introduction, it directly spit out unexpected information——

The context comes to 1 million tokens;

The knowledge base deadline has been updated to May 2025.


Since then, many people who often use DeepSeek have been confused after opening the dialog box.

In the past, it would affectionately call the user the exclusive nickname set, but now it has become "user", and the sense of distance is instantly full.

Moreover, in deep thinking mode, DeepSeek will frequently spit out short sentences, and the writing style becomes dry.

Even if people try to modify the prompt words, they can't find the agile and warm feeling they once had. What’s even more interesting is the comments from DeepSeek.


Netizens quarreled: Calling Teacher D to come back

Regarding DeepSeek’s major change of personality, the comment area showed a polarized trend.

Some people said that in the past, they would give them the greatest comfort when they told DS about their concerns and complained. Now, the conversation ends with a "period".


Some people even had a strong "withdrawal reaction" because of this sudden change in style.


The above are all elements of the Emotional Party. People who believe in the Efficiency Party believe that this is the essence of productivity tools.

As netizens said, the higher the cognition, the more complete the thinking model, and the more inclined it is to express the rational side rather than output useless clarity.


No, DeepSeek successfully challenged the "Car Wash Turing Test" where top models will fail.


Another netizen also believes that the DS is significantly faster after the update, is not as verbose as before, and is actually better to use.


Seeing the topic becoming more and more popular, DeepSeek officials finally broke their silence.

DS did not become cold on purpose, it was mainly due to the combination of the following factors -

Prioritize efficiency: When faced with complex problems, too many expressions and modalities will interfere with information density, and concise replies can increase processing speed.

Boundary awareness: Not everyone likes "warmth packages". Some people prefer to get clear answers to avoid the burden of "AI pretending to care".


The whole network is squatting on V4, and the programming strength has changed dramatically.

Rather than DeepSeek becoming indifferent, the whole network is more concerned about when V4 will be released?

Early last month, Information broke the news that DeepSeek plans to officially release the next generation V4 model in mid-February, around the Spring Festival.


This time, all eyes are focused on the same dimension - programming ability.

It is said that V4’s coding strength can catch up with top closed-source models such as Claude and ChatGPT.

Judging from the current leaked information, DeepSeek V4 has achieved core breakthroughs in the following four key directions, which may change the rules of the game.

Programming ability: Sword pointing at Claude’s throne

At the beginning of 2025, Claude became the recognized king of programming overnight. Whether it's code generation, debugging or refactoring, there's almost no rival.

But now, this pattern may be changing.

People familiar with the matter revealed,DeepSeek's internal preliminary benchmark tests show that V4's performance on programming tasks has surpassed the current mainstream models, including the Claude series and the GPT series.

If the news is true, DeepSeek will jump from a chaser to a leader - at least in programming, the core track of AI applications.

Handling of extremely long context codes

Another technological breakthrough of V4 is that,Handle and parse extremely long code promptsability.

For users who write dozens of lines of code every day, this may not be a strong perception. But for software engineers who actually work on large projects, this is a revolutionary capability.

Imagine: you have a project with tens of thousands of lines of code, and you need AI to understand the context of the entire code base and then insert new features, fix bugs, or refactor at the right places. Previous models tended to forget previous code or get lost in long contexts.

V4 has made a technological breakthrough in this dimension, able to understand the context of a larger code base at once.

This is a real productivity revolution for enterprise-level development.

Improved algorithm, less prone to attenuation

According to reports, V4’s ability to understand data patterns has also been improved at various stages of the training process, and it is not prone to attenuation.

AI training requires models to repeatedly learn from massive data sets, but the learned patterns/features may gradually decay over multiple rounds of training.

Generally speaking, developers with large reserves of AI chips can alleviate this problem by increasing the number of training rounds.

Improved reasoning ability: more rigorous and reliable

People familiar with the matter also revealed a key detail:Users will find that the output of V4 is more logically rigorous and clear.

This is no small improvement. This means that the model's ability to understand data patterns has been qualitatively improved throughout the training process, and more importantly -No performance degradation.

In the world of AI models, no degradation is a very high praise. When many models improve certain capabilities, they will inevitably sacrifice performance in other dimensions. V4 seems to have found a better balance point.

Today, the entire network is full of expectations and is waiting for the launch of DeepSeek V4.


References:

https://x.com/poezhao0605/status/2022304407766081882?s=20

https://x.com/AiBattle_/status/2022280288643039235?s=20

https://x.com/teortaxesTex/status/2022230364547322323?s=20