According to media reports citing multiple informed sources, DeepSeek founder Liang Wenfeng recently indicated in internal communications that the company's new flagship large model, DeepSeek V4, will be officially released in late April. It is understood that DeepSeek V4 will not only feature a trillion-parameter scale and a million-level context window but will also achieve, for the first time, deep compatibility with domestic chips such as Huawei's Ascend. This development marks a critical step forward for China's AI industry on the path toward "de-CUDAization."
On April 8, the DeepSeek web interface quietly introduced two new interaction modes: "Fast Mode" and "Expert Mode." These modes support simple searches as well as the processing of long-range and complex queries, sparking industry speculation about the imminent release of the V4 model. This is the first time since DeepSeek's rise in popularity that such a tiered design has been implemented at the product level. Fast Mode is suited for daily conversations, offering quick response times and the ability to directly recognize text within images and documents. Expert Mode, meanwhile, is more adept at handling complex tasks, supporting deep thinking and intelligent search functionalities, although it currently does not support file uploads or multimodal capabilities.
Analysts suggest that the introduction of these two interaction modes strongly indicates that DeepSeek is preparing for a more complete model matrix. In the future, the emergence of different entry points—potentially lighter, more in-depth, or even more multimodal—cannot be ruled out.
In a related development, according to media reports, in preparation for the DeepSeek V4 launch, major domestic technology giants including Alibaba, ByteDance, and Tencent have already placed orders for several hundred thousand units of a new generation of AI computing chips. These companies plan to offer the new DeepSeek model via cloud services and integrate it into their own AI products.
Comments