Chinese start-up DeepSeek unveiled DeepSeek-V4 on Friday, claiming "significantly enhanced" capabilities versus its predecessor R1 — the January 2025 model that briefly shook Wall Street by proving that a cheap Chinese AI agent could credibly compete with ChatGPT. The announcement lands the same week OpenAI presented GPT-5.5, and while US authorities accuse DeepSeek of using restricted Nvidia Blackwell chips for training. According to The Information, V4 was actually optimised to run on Huawei chips rather than US silicon.
The Hangzhou-based firm released two variants: V4-pro, the more capable and compute-hungry version, and V4-Flash, a cheaper and lighter model. Neither is multimodal yet (no video/image processing), but DeepSeek said it is working on adding those capabilities. This is the headline technical limitation versus frontier Western models.
The launch matters because China's AI sector has been under intense scrutiny. US export controls have starved Chinese labs of the best Nvidia chips, and Moonshot AI, Zhipu AI and MiniMax were seen as getting ahead. DeepSeek also lost key talent — Guo Daya, a lead architect of the R1 model, was reportedly poached by ByteDance at over 100 million yuan (12.5 million euros). V4 is DeepSeek's answer that it is still in the game. Source: Les Echos, 24 April 2026, Claude Fouquet.