R1 core author departs but the team remains intact, V4 possibly released in April: DeepSeek's next step is Agent

According to 1M AI News monitoring, LatePost reported that DeepSeek V4 may be released in April. A small-parameter version was delivered around January this year to some parts of the open-source framework community for adaptation; the large-parameter version had previously been expected to be released around mid-February, but it was delayed until now. LatePost believes V4 is likely still the strongest open-source model, but it will be difficult to outclass other top-tier strong models, because the criteria for what “strong” means have become increasingly diverse for developers and users across different scenarios. And after entering the Agent era, product reach and long-tail usage data have become even more important—this is precisely the area where DeepSeek had not put much effort before.

Since the second half of 2025 up to now, four core members have clearly left:

1. Wang Bingxuan, the core author of DeepSeek’s first-generation large language model; later participated in training for successive models; at the end of last year, he was recruited by Tencent’s Yao Shunyu
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; formally left recently
4. Ruan Chong, a key multi-modal contributor such as Janus-Pro; joined the autonomous driving company Yuanrong Qixing in January this year

LatePost said the team did not see organized departures. Competitors have offered total packages that are 2–3 times higher, or even in the eight-figure range, leading more people to choose to stay. DeepSeek has not raised funding to date and has no clear valuation. In 2023, founder Liang Wenfeng had met investors on a small scale and proposed repayment-maximum conditions similar to the return cap terms in OpenAI and Microsoft’s investment agreement; no institutions accepted, and after that, he stopped meeting investors. MiniMax and Zhipu have gone public in succession and their stock prices have surged; employees’ questions about unpriced options held in their hands have increased, and Liang Wenfeng has recently started trying to come up with a way to value the company.

Signals of a shift in product direction have appeared. In a hiring post released by a DeepSeek HR in mid-March, the first time it mentioned specific product names, it required a candidate for a “model strategy product manager” in the Agent direction who is “familiar with and has deeply used well-known agents such as Claude Code, OpenClaw, and Manus.” DeepSeek already has a product team of a few dozen people, but previously it had not ventured into the AI programming and general Agent direction; for C-end, it still only has Chatbot.