R1 Core Developer Resigns but Team Stays Together, V4 to be Released in April: DeepSeek's Next Step is Agent

According to 1M AI News's monitoring, as reported by LatePost, DeepSeek V4 is likely to be released in April. A minor parameter version was already handed over to some open-source framework communities for adaptation around January of this year, while the major parameter version was previously expected to be released around mid-February, but has been delayed until now. LatePost's assessment indicates that V4 is still probably the strongest open-source model, but it is difficult to be overwhelmingly strong, as developers and users in different scenarios have increasingly diverse standards for "strength." Moreover, since entering the Agent era, product outreach and long-tail usage data have become more important, an area where DeepSeek has not invested much in before.

From the second half of 2025 to the present, four core members have clearly departed:

1. Wang Bingxuan, the original core author of DeepSeek's large language model, who has been involved in training successive models and was poached by Tencent's Yao Shunyu at the end of last year.
2. Wei Haoran, the core author of the DeepSeek-OCR series, left around the time of the Spring Festival.
3. Guo Daya, the core author of DeepSeek-R1, officially resigned recently.
4. Ruan Chong, a core contributor to Janus-Pro and other multimodal projects, joined an autonomous driving company, Element Rongqi, in January of this year.

LatePost stated that the team did not suffer a mass exodus. Competitors have offered total compensation packages 2-3 times or even 8 figures higher than DeepSeek's offers, but more people have chosen to stay. DeepSeek has not raised any funds to date and does not have a clear valuation. In 2023, founder Liang Wenfeng had limited interactions with investors, proposing a return cap condition similar to the OpenAI and Microsoft investment agreement, which no institution accepted. After that, there were no more investor meetings. The IPOs of MiniMax and Zhipu have seen their stock prices soar, leading to more questions from employees about the unpriced options in their hands. Recently, Liang Wenfeng has begun considering ways to value the company.

There are signals of a shift in product direction. In mid-March, a DeepSeek HR employee mentioned a specific product name for the first time in a recruitment announcement, seeking a "Model Strategy Product Manager" for the Agent direction who is "familiar with and has extensively used well-known agents like Claude Code, OpenClaw, Manus." DeepSeek already has a product team of several tens of people, but has not previously ventured into AI programming and general Agent directions, with the consumer end still limited to Chatbot.