Author: Zhang Chi, source: "Chuangzhi Ji" by Phoenix Technology

After several rumors of "bounced tickets", the much-anticipated DeepSeek V4 is about to debut.

According to "Sharing Intelligence", we learned from multiple people familiar with the matter that DeepSeek founder Liang Wenfeng recently revealed in internal communications that DeepSeek's new generation flagship large model DeepSeek V4 will be officially released in late April.

Previous relevant news showed that DeepSeek V4 will not only bring trillions of parameter scale and million-level context windows, but also achieve in-depth adaptation with domestic chips such as Huawei Ascend for the first time, marking a key step for China's AI industry on the road to "de-CUDA".

It is worth noting that on the evening of March 29, DeepSeek experienced a 13-hour service outage. At that time, the official only called it "server failure", but after recovery, many developers found that the model output logic had changed significantly: the coding structure was clearer, the logic rigor was greatly improved, and the SVG graphics generation capability was significantly enhanced.

It is widely speculated in the industry that this "suspension" is actually a grayscale test or full deployment of V4.

It is reported that V4 will be open source under the Apache 2.0 protocol, and enterprises can deploy it independently. According to the latest foreign media reports, in preparation for the release of V4, domestic technology giants such as Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of new-generation AI computing chips, and plan to provide the new DeepSeek model through cloud services and integrate it into their own AI products. Affected by this, the price of new AI chip products has increased by about 20% recently.

Previously, the release date of DeepSeek has been postponed many times. It was originally planned to launch the large-parameter version around the Spring Festival in mid-February, but it was later postponed due to in-depth adaptation of domestic computing power, architecture reconstruction, and stability polishing.

Liang Wenfeng, who has invested in cross-border AI from quantitative investment, is experiencing his own "transformation".

"Late" reported that in the past year, although some core members of the DeepSeek team have resigned (including R1 core author Guo Daya, etc.), "some people left, but more people stayed", and the team has not become an organic loss.

Since the second half of 2025, Liang Wenfeng has begun to mention productization and commercialization more, forming a product team, recruiting "model strategy product managers" in the Agent direction, and extending from pure model research and development to the application layer. At the same time, he is valuing the company to clarify the value of employee options and give the team more certain expectations.

Facing strong competition from rivals such as Doubao, Qianwen, Zhipu, Kimi, and MiniMax, V4, which was delayed for several months, carries huge pressure of expectations - can Liang Wenfeng once again create the world's most outstanding open source large model?

It is understood that the biggest technical highlight of V4 is the breakthrough in LTM (Long Term Memory) capabilities.

However, many AI entrepreneurs who have in-depth contact and cooperation with DeepSeek told "Shark Tank" that the industry should not have too high expectations. It is difficult for V4 to replicate the influence of last year's DeepSeek-V3.

In addition to DeepSeek V4, the Hunyuan 3.0 large model will also be officially released in April. This is the first major model product launched by Tencent since it completed the reorganization of its team structure in the second half of 2025 and Yao Shunyu became the chief AI scientist.