The Whole Village Is Waiting for DeepSeek to Serve Dinner

Anticipation meets silence

It has been reported that the Financial Times and other outlets said DeepSeek will unveil “V4” on March 2 — and that V4 will be optimized for domestic compute. Reportedly this would make V4 the first in its series to run fully on China’s own AI chip and cloud stack. The claim set off a wave of domestic coverage and public expectation. Then nothing happened.

What’s actually new — and what’s not

DeepSeek (a Chinese AI startup widely known by its English name) has been iterating: a recent update widened the context window and was widely mistaken for a V4 beta, and the team’s papers — many authored by Liang Wenfeng — show active work on code generation and multimodal capabilities. Why those tracks? Because AI coding is the fastest-maturing use case, and multimodality is the company’s obvious weak spot. But it has been reported that none of the media accounts about V4’s training pipeline or infrastructure have been confirmed by the company; publicly available information is mostly academic, not engineering telemetry.

Engineering realities and geopolitical friction

Switching a major model from one compute stack to another is not just paperwork. Even within Nvidia’s ecosystem, moving architectures takes time; between entirely different domestic and foreign frameworks the engineering burden grows dramatically. If V4 truly embraces domestic chips and clouds, the transition would carry both technical risk and geopolitical implications amid trade and export controls affecting AI hardware. Reportedly, that is one reason why the team may be pacing the rollout rather than racing to meet competitors.

Stakes and timing

Despite a year without a headline upgrade, DeepSeek’s app reportedly still exceeds 100 million MAU and ranks among China’s top AI applications. Competitors such as Doubao (豆包), Qianwen (千问) and GLM rushed updates over the Spring Festival in what looked like pre-emptive moves. So when will V4 “serve dinner”? The market knows the release is near — but whether it arrives as a standard V4, a parallel deep-inference R2, or a combined launch will determine whether DeepSeek reclaims global attention or cedes ground in an ever-accelerating AI arms race.