DeepSeek V4 Returns After a Year in Hiding, Challenging Nvidia's Dominance and Issuing a Warning to OpenAI

Token consumption in China has surged more than 1,000-fold over the past two years, and DeepSeek is preparing a new challenge to OpenAI as “agent” applications accelerate demand for model usage. After more than a year of limited public activity following DeepSeek V3 and R1, expectations for a DeepSeek V4 release have resurfaced, alongside signals of system incidents and the recent appearance of an “expert mode.”

The Chinese AI story has shifted

During the 2025 Lunar New Year, DeepSeek R1 launched with low costs, strong performance, and open-source code, quickly drawing global attention. Over the following year, China’s AI landscape evolved into a multi-player environment rather than a single race led by DeepSeek.

According to the article, the market has formed a “(3+1) + 6 + N” structure:

• (3+1) giants: ByteDance, Alibaba, Tencent, and Baidu
• “6” group: Kimi, Zhipu AI, MiniMax, StepFun, Baichuan, and ModelBest, with several either already public or planning IPOs
• “N” vertical/niche firms: specialized AI companies that DeepSeek must overcome to regain its position

Tokenomics and the rise in usage

The article frames tokenomics as the “value loop” of the intelligent economy. It contrasts earlier industrial-era measurement of data flow (GB) with the AI-era focus on tokens as a measure of capability. Token consumption in China has risen to 140 trillion calls per day by March 2026, though the article notes that tokens measure quantity rather than quality.

It also cites a view attributed to La Phuc Li from Xiaomi, who said that global compute supply cannot keep up with token demand driven by agents. The “escape,” the quote continues, is not simply cheaper tokens, but “co-evolution”—coordination between a token-efficient agent framework and more powerful, efficient models.

DeepSeek V4: expected capabilities and efficiency focus

The article says expectations for DeepSeek V4 are not centered on a simple parameter increase. Based on scientific papers cited by Huxiu, DeepSeek V4 is described as integrating multiple architectural breakthroughs:

• OCR 2 and causal vision: enabling the model to read multi-million-page documents in a human-like logical order, aimed at improving handling of complex documentation
• mHC: designed to address information bursts in billion-parameter-scale models, supporting stable training without relying on the most advanced process nodes
• Engram and long-term memory: described as redefining the Transformer architecture by preserving state across interactions at low cost
• Agent software optimization: aimed at doubling inference efficiency for agents, addressing GPU underutilization

What the next model cycle could mean

The article suggests DeepSeek’s upcoming flagship model may integrate text, image, and video generation, with a trillion-parameter scale and features such as memory, tool use, strong learning, and support for agent-based workflows. It also highlights an objective to adapt to domestic compute capabilities in China, including reports that DeepSeek has prioritized early access for domestic chipmakers such as Huawei to optimize performance.

If DeepSeek V4 can run stably and deliver top-tier performance on domestic chips, the article says it could weaken Nvidia’s monopoly and strengthen China’s technology security. It also argues DeepSeek would need to lead in adopting a new scaling approach—achieving prior large-model capability with smaller-scale models—and support “Sugar Intelligence” (1B–8B models running smoothly on phones) to broaden access, while also pursuing “Force Intelligence” (trillion-parameter models) for high-value tasks such as scientific research and finance.

From model competition to ecosystem competition

The article characterizes 2026 as a turning point from “model battle” to “ecosystem battle,” emphasizing the need for a sovereign intelligent operating system rather than only a more capable chatbot. It concludes that DeepSeek’s prolonged silence may have been preparation for a resurgence, with V4—and potentially a future R2—positioned as more than a product update for China’s AI standing.

Source: Huxiu