The whole village is waiting for DeepSeek to come to the table for dinner.

Last weekend, some foreign media, including the UK's Financial Times, reported that DeepSeek V4 would be released this Monday (March 2nd). Other reports claimed that V4 would be optimized for domestic chips and would be the first in the series of large models to be fully based on the domestic computing power ecosystem. This news was quickly reprinted by many domestic media, and everyone was looking forward to seeing V4 unveiled - but nothing happened.

People have been looking forward to the release of DeepSeek V4 for over a month. Previously, many people predicted that it would be released during the Spring Festival, but there were also reports saying that the Spring Festival was too early. In fact, during the Spring Festival, competitors like Doubao, Qianwen, and GLM rushed to release new versions, which can be seen to some extent as a "preventive competitive behavior" against DeepSeek: once V4 is released, its brilliance will be so overwhelming that it will outshine all competing products; so competitors must also quickly come up with high - level new versions.

Shortly before the Spring Festival, DeepSeek made an update, expanding the context window, and it was once misreported that "V4 had started gray - scale testing". During that time, people in my WeChat circle were constantly asking, "Have you been included in the V4 gray - scale test? How can I get an invitation code?" As it turned out, that update was not V4. Although it was widely called "V4Lite", it was not the official version of V4; its benchmark scores placed it in the first echelon of domestic large models, but the update was not eye - catching enough.

My friends in the AI circle generally still have high hopes for DeepSeek V4, and there are two reasons for this:

Judging from a series of papers signed by Liang Wenfeng, the DeepSeek team has ideas in basic research and is trying to innovate. Its stance is higher than that of most large - model startup teams;

At the end of January this year, Liang Wenfeng confidently previewed an upcoming new version, "Given his personality, he won't promote something without being sure of it."

Judging from the papers published by the DeepSeek team, its research and development in recent times seems to focus on two directions: one is programming, and the other is multimodality. The former is very natural because AI programming is the fastest - growing and most mature large - model application. Since the beginning of this year, Claude Code with Opus 4.6 and ChatGPT - 5.3 - Codex have once again raised the bar for AI programming. This is a proven path, and the previous minor versions of DeepSeek have also made great progress in this area. It is understandable that V4 will make significant progress again.

The latter is to address a weakness - DeepSeek's biggest weakness so far is the lack of multimodal functions. This not only limits its C - end applications but also its B - end applications. Since the source of B - end revenue is the customer's consumption of Tokens, the Token consumption of multimodal applications is one or even several orders of magnitude higher than that of traditional text generation. The recent release of Seedance 2.0 once again proves the high user base and commercial potential of excellent multimodal large models. Judging from the published papers, DeepSeek is definitely not satisfied with being just a "text - to - text" large model and has put a lot of effort into multimodality.

Now, everyone is waiting for DeepSeek V4 to be launched. When will it really happen? I think there is a very important factor: V3/R1 was the world's most outstanding open - source large model at that time, especially making great progress in inference cost control, which shocked the world; at this moment, V4 must also become the world's most outstanding open - source large model (at least one of them) to shock the world again. This is a problem that successful people must face: the benchmark of success keeps rising, and each challenge is greater.

There is another important factor: what shocked the world at that time was mainly R1, the deep - inference large model. Without the deep - inference function, at least for ordinary users, DeepSeek would not seem so special. Suppose DeepSeek first launches the "ordinary version" of V4 and then launches the "deep - inference" R2 after some time. Will the market be satisfied? Of course, a more conservative approach is to launch V4 and R2 simultaneously, but this requires more resources, which may not be suitable for a startup - level company.

There is also a problem: Large - model development is a competition in both basic research and engineering execution. The public information we can currently see is mainly at the basic research level, such as the academic papers published by the DeepSeek team; we know very little about the engineering issues and bottlenecks behind it. So far, all media reports about the training process of the new version of DeepSeek are actually just speculation and have not been officially recognized, nor do they have third - party reliable sources. For example, Google has always been leading in basic research on large models, but it made many mistakes in engineering execution in the early stages of Bard and Gemini and only caught up in the second half of 2024.

If DeepSeek V4 really fully embraces the domestic computing power system as reported by the UK's Financial Times, the engineering challenges will be even greater. You know, even within the NVIDIA framework, it takes some time for Silicon Valley tech giants to shift training tasks from the Hopper architecture to the Blackwell architecture; let alone adjusting between two completely different frameworks. Such engineering problems are destined to be difficult to solve in the short term. It's good enough if they can be solved, and we can't demand too much.

However, it should be emphasized that the current reports on the training details of V4, whether from foreign or domestic media, are basically speculations and integrations from indirect channels and have never been officially recognized. The real training details will only be known to the outside world when the new version is released, and not all details will be revealed. We can only say: Anything is possible, but not every possibility is worth discussing now.

By the way, even without a major version update for over a year (with 2 - 3 minor version updates in between) and with very few marketing activities, the MAU of the DeepSeek APP still exceeded 100 million, ranking fourth or fifth among domestic AI applications. Even Yuanbao, which spent a lot of money during the Spring Festival, couldn't surpass it (to some extent, this is thanks to the V4Lite update). If V4 is launched now, DeepSeek still has a chance to rush into the top three in the domestic AI field; if it's launched later, it's hard to say, because the arms race among Internet giants in the AI field is intensifying, and the competition density that DeepSeek faces this year is much higher than last year, both at home and abroad.

Anyway, since there have been frequent reports in the market that "DeepSeek V4 is about to be released", it at least shows that the release of the new version is not far off. I really hope to see V4 soon because so far, DeepSeek is still one of the four large models I use most frequently (the other three are GPT, Gemini, and Grok), and it is also the domestic large model I use most often. I will definitely try it as soon as it is released and ask my friends in the AI circle about their experiences. I hope we won't be disappointed!

This article has not been funded or endorsed by DeepSeek or any of its competitors.

This article is from the WeChat official account "Internet Phantom Thief Group" (ID: TMTphantom), written by Pei Pei, the leader of the Phantom Thief Group, and is published by 36Kr with authorization.