This Spring Festival, the entire Internet is waiting for DeepSeek to release the latest large model DeepSeek V4, but now it seems that the large version may not be released during the Spring Festival. Today, DeepSeek quietly updated the new version, which has many highlights. According to DeepSeek's own introduction, this update mainly improves the context capability, reaching 1M, while the previous DeepSee V3 series was 128K.
With 1M context, DeepSeek can process extremely long text of the size of the "Three-Body Problem" trilogy at once.This has been confirmed by actual tests by many netizens.
Other capabilities have been available before, and DeepSeek did not mention more detailed information.
It is worth noting that DeepSeek added a special paragraph after answering this question, emphasizing that it is the latest DeepSeek model, without a specific subname, not V3 or R1, as if to deliberately emphasize the difference of this model, but not to make people think that this is the V4 Preview version.
However, judging from the feedback from netizens, this small update of DeepSeek has actually greatly improved its capabilities. The well-known large model testing blogger @karminski-dentist has released a quick test, which mainly tests an elephant toothpaste test project in terms of programming.
The conclusion is that visible aesthetics, modeling, physical simulation, and code performance have all been improved, and there are even small surprises.This is SOTA than any version released before.That is, the current optimal large model (of course, mainly refers to the DS version).
It is not yet certain whether DeepSeek's new model is complete, because some netizens tested the question of whether to walk or drive to a car wash at a distance of 50 meters, which has been very popular these two days. DeepSeek's new model gave different answers. Some said it was correct, and some said it was not possible.