DeepSeek officially announced that it is testing a new large model, which may be V4 Lite

There has been an explosion of domestic AI in the past few days. GLM-5, Minimax 2.5 and DeepSeek all released new large models on the same day on the 11th, among which DeepSeek's naturally attracted the most attention. We have reported before that this update mainly improves the context capability, reaching 1M, while the previous DeepSee V3 series was 128K, which is 7 times higher than the previous V3 series large model.

DeepSeek also officially confirmed this in the official group tonight,Indicates that the web page and APP version are testing a new long text model structure and support 1M context.

At the same time, DeepSeek also emphasized that the API service has not changed. It is still a V3.2 series large model and only supports 128K context.

Judging from the introduction of DeepSeek, this new model is still a text model. The main improvement is contextual capability, which is also very important in many fields. During long conversations, it is easy for large models to be unable to remember previous content due to insufficient context.

Although there have been many actual tests on the Internet showing that this DeepSeek large model has greatly improved in terms of programming, output speed, etc., but compared to previous expectations, this update is inevitably a bit disappointing.

The large model this time is obviously not V4, but more likely V4 Lite, because the number of parameters is reported to be only 200 billion, which is much less than the 670 billion of the V3 series, so it is normal for some capabilities to be worse than V3.

It is speculated that this model is V4 lite. It is unlikely that DeepSeek will release only one large V4 model in the future. Instead, there will be different versions. Each series has different directions and designs. The current V4 Lite is just a pathfinder, so there is not much improvement. Moreover, DeepSeek officials have not detailed its technical architecture, and more information has yet to be released.

The rumored DeepSeek V4 full-blooded version has 1.5 trillion parameters, more than double the V3 series.New technologies such as Engram and mHC previously studied by DeepSeek will also be used.The performance is comprehensively improved while the cost is still low. This expectation is still very high.