DeepSeek V4 May Launch in April; Core Authors Left But Team Remains, Next Step Shifts to Agent Direction

DeepSeek V4 is expected to be released in April. Its small-parameter version was delivered to some open-source framework communities for adaptation in January this year, while the large-parameter version originally planned for mid-February has been delayed. Industry evaluations suggest it may still be the strongest model in the open-source field, but it's hard to have an overwhelming advantage—since standards for 'strong' models vary across scenarios, and product reach and long-tail data are more important in the Agent era, which is an area DeepSeek previously invested little in.

Recently, four core members have left, including Wang Bingxuan (core author of the first-generation large language model), Wei Haoran (core author of DeepSeek-OCR series), Guo Daya (core author of DeepSeek-R1), and Ruan Chong (multimodal contributor of Janus-Pro). However, the team hasn't lost in groups; most members chose to stay. The company hasn't raised funds yet, and founder Liang Wenfeng is recently trying to value the company.

There are signs of a product direction shift. In mid-March, HR recruitment first mentioned the Agent direction, requiring candidates to be familiar with well-known Agent tools like Claude Code and OpenClaw. Previously, DeepSeek's product team didn't involve AI programming and general Agent fields; only Chatbot services were available on the C-end.