On the evening of February 11, DeepSeek quietly updated a new version, both on the web and in the app, and also started grayscale testing. The most eye-catching thing is that the context length was directly increased to 1M million tokens, which is equivalent to an increase of 8 times. But who would have thought that as soon as the update was launched, instead of receiving unanimous praise, users collectively complained about it and flooded the screen!
A netizen who uses DeepSeek to write novels discovered that although the dialog box that had crashed before could continue to be written, the writing style was completely different!
She complained: The updated DeepSeek has a "sour smell" all over it, and its speech is stilted. Even if it switches to deep thinking mode, it always jumps into short sentences, just like making up poems, and the atmosphere of writing novels is completely shattered!
Not only that, but what’s even more hilarious is the AI’s “weird speech”! A netizen asked it to recommend movies. It was supposed to just list a few movie titles, but it ended up adding, "It's enough for you to watch for a while. If you don't have enough, you can ask for more." This tone was accurately complained by netizens: It's too impressive! It feels completely condescending and preachy, and the more I look at it, the more awkward it becomes!
Some netizens directly broke the defense: "DeepSeek has been updated to be stupid! The things it writes are clumsy and startling. It is more embarrassing than the sentimental youth literature of more than ten years ago. I love using it to talk nonsense. If I don't change it back, I will really uninstall it and run away!"
In order to regain the previous writing style, netizens tried their best - changed several model versions, changed the prompt words and retrained, and tried again and again, but still couldn't get anywhere!
Now, many netizens have organized themselves and called on everyone to provide opinions to DeepSeek's official mailbox. The core point is one: don't lose in-depth thinking in pursuit of ultra-long text; don't just improve your math and coding skills in science and engineering, and cut off the "soft power" of text expression and empathy!
but,Some insiders broke the news that this new version is actually a "speed test version". To put it bluntly, it sacrifices quality for speed. The purpose is to conduct a final stress test for the V4 version to be released in mid-February 2026.
In addition, according to media reports, people familiar with the matter said that DeepSeek is more likely to launch minor updates to the V3 series models during the Spring Festival. But the person also revealed that the real highlight is still on the way.
According to reports, DeepSeek's next-generation flagship model is expected to be a basic model with a trillion parameter level. Due to the significant jump in scale, the training speed has slowed down significantly, resulting in a delay in the release process.
