Inside a nondescript office tower on the western edge of this tech hub, a handful of engineers are putting the finishing touches on a piece of software that has quietly become the most anticipated test of China’s artificial intelligence ambitions in 2025.
The software belongs to DeepSeek, a two-year-old startup spun out of the quantitative trading firm High-Flyer. Until late last year, DeepSeek was known mainly among AI researchers for releasing efficient, open-source models that performed respectably against American rivals. Then, in December 2024, DeepSeek launched DeepSeek-V3 a model that stunned Silicon Valley by matching GPT-4’s reasoning on several benchmarks while costing less than $6 million to train, a fraction of the hundreds of millions spent by OpenAI or Google.
Now the same team is preparing an even more powerful successor, tentatively referred to internally as DeepSeek-V4 or “Nova.” The new model is expected to debut within weeks, according to two people familiar with the company’s roadmap who spoke on condition of anonymity because the launch date remains confidential.
For Beijing, the stakes could not be higher. The United States has spent the past two years tightening export controls on advanced semiconductors, aiming to slow China’s progress in frontier AI. NVIDIA’s H100 and Blackwell chips are off-limits; even the restricted H800 has been blocked. Chinese firms have been forced to work with less powerful alternatives or stockpiled hardware.
DeepSeek became a poster child for overcoming those restrictions. Its V3 model was trained using a cluster of roughly 2,000 Nvidia H800 chips, powerful but not the latest combined with innovative software tricks that squeezed out maximum efficiency. The company published its technical papers in full, a rarity among Chinese AI labs, earning goodwill from global researchers.
“DeepSeek showed that clever engineering can partly offset hardware bans,” said Wei Liang, a Singapore-based AI analyst who has followed the company closely. “But a new model will have to prove that wasn’t a one-time miracle. Investors and policymakers want to know if China can sustain that pace without access to cutting-edge chips.”
The new model is expected to feature a “mixture-of-experts” architecture, similar to V3 but with a larger number of specialized sub-models. Early indicators from internal tests, described in Chinese social media posts from researchers who have seen partial results, suggest Nova scores within 5% of GPT-4 Turbo on coding and mathematical reasoning tasks, a narrow gap that would mark a significant achievement given the hardware disadvantages.
Yet doubts linger. Some Western AI researchers have privately questioned whether DeepSeek’s reported training costs omit certain expenses, such as prior research and failed experiments. Others note that running the model at scale remains expensive, even if training was cheap. And no one outside DeepSeek has independently verified the company’s performance claims on flagship benchmarks like MMLU or HumanEval.
DeepSeek declined to comment for this story. But its founder, Liang Wenfeng, hinted at the company’s trajectory in a rare October interview with a Chinese tech publication. “Efficiency is our weapon,” Liang said. “We are not trying to build the largest model. We are trying to build the smartest model for the resources available.”
That philosophy aligns with a broader shift in China’s AI strategy. After the initial shock of U.S. chip sanctions, the country’s tech giants Baidu, Alibaba, Tencent, and ByteDance have focused on optimizing existing architectures rather than chasing parameter counts. DeepSeek, unburdened by legacy products or cloud customers, has moved fastest.
But the new model will face immediate competition. Alibaba’s Qwen team is expected to release Qwen-3 later this spring. Zhipu AI, backed by government funds, recently previewed its GLM-5 model. And abroad, Anthropic and OpenAI continue to push the frontier with models like Claude 4 and GPT-5, both rumored for late-2025 releases.
For now, the AI world watches Hangzhou. DeepSeek’s next release will not just determine the startup’s fate. It will signal whether China’s AI industry can truly decouple from U.S. chips and still compete at the highest level or whether the sanctions have drawn a permanent ceiling.
“DeepSeek is the canary in the coal mine,” said Wei Liang. “If Nova performs close to state-of-the-art, expect a wave of investment and a recalibration of U.S. export controls. If it stumbles, the narrative shifts to a long, slow decline.”
The company has scheduled no press conference. No keynote. No celebrity CEO on stage. Just a GitHub repository and a technical paper, ready to go live on a quiet Tuesday morning. Then the world will know.
