Rumors Suggest DeepSeek's Next Major Model Is Coming: 3 Trillion Parameters, Performance Aiming for GPT-6 Astra

This week, DeepSeek released V4.1 Flash. The changes in this large model architecture are so significant that it can be considered as V5, but DeepSeek remained low-key and did not change the major version number. The new product of the Pro series has been confirmed as V4.1 Pro, but specific upgrades have not been announced yet; if the improvement from V4 to V4.1 is similar, the performance ceiling of V4.1 Pro will be very high, but considering that the previous V4 Pro-0813 final version did not meet expectations, the external expectations should not be raised too high.

DeepSeek Code 2.0 Is Said to Be About to Launch, Focused on Computer Control

According to leaks, there will be another big model named DeepSeek Code 2.0, which is expected to be released in September, and its performance is expected to surpass Mythos 5.1 and GPT-6 Astra—currently the two most powerful cutting-edge large models.

In other aspects, DeepSeek Code 2.0 is expected to continue maintaining open source and openness, with a design focus on Computer Use, meaning letting AI replace users in operating computers to complete a large amount of work. Many impressive features of Astra come from this capability.

3 Trillion Parameters, the Largest Scale in China

It is reported that the model has more than 3 trillion parameters, which is obviously also the key to improving its performance. By comparison, the current strongest in China is K3's 2.8 trillion parameters, Qwen 3.8 Max has 2.4 trillion parameters, and DeepSeek V4 Pro has 1.6 trillion parameters.

Considering that V4.1 Flash changed to a new structure, doubling the parameter count, the credibility of DeepSeek Code 2.0 increasing from the current 1.6 trillion parameters to 3 trillion is there—competing with large models like Mythos 5.1 and Astra requires such a scale of parameters, otherwise it would be difficult to improve the performance ceiling.

However, the credibility of this leak is still hard to judge. DeepSeek released three large models named Code between 2023 and 2024, and they performed rather modestly at that time; this time, if it re-launches a code/logic-oriented Code large model as a flagship product, it would be a reasonable direction, leaving V4.1 Pro and Flash to handle general agent tasks.