Chinese AI developer DeepSeek has released a preview of its next major model, V4, with open weights and a particular emphasis on writing computer code. The release, reported by Caixin and others in the last week of April, is the latest move in a strategy that has turned open source models into one of China’s most effective tools in the global contest over artificial intelligence.
The “DeepSeek moment” and what followed
DeepSeek drew worldwide attention in early 2025 when it published details of a model that it said had been trained at a fraction of the cost of comparable systems from OpenAI and Anthropic, using far fewer of the expensive chips that dominate AI budgets. The episode, widely called the “DeepSeek moment,” rattled technology stocks and prompted a reassessment of how far behind Chinese labs really were.
Just as important was how the company distributed its work. DeepSeek released its models openly, allowing developers anywhere to download, run and modify them. OpenAI and Anthropic, by contrast, keep their most capable models behind paid interfaces. Within months, other Chinese companies including Alibaba and Moonshot AI followed with dozens of open models of their own.
What V4 adds
According to the release notes summarized by NYU Shanghai’s research technology group, the V4 preview comes in two versions, a larger “Pro” model and a smaller “Flash” model, both using a mixture-of-experts design and supporting a context window of about one million tokens. DeepSeek says the models were designed to run efficiently and that the preview performs strongly on coding benchmarks, where it leads other openly available models by the company’s measures. Independent evaluations are still emerging, and the leading American closed models retain an edge on most published tests.
Moonshot AI released an updated open Kimi model in the same period, intensifying competition among Chinese labs themselves.
Coding ability matters because it is the foundation for AI agents, systems that can operate office software such as spreadsheets, email and calendars with limited supervision. The same skills make models better at finding security flaws in software, a capability that cuts both ways for defenders and attackers.
Distillation disputes and export controls
The rise of Chinese open models has not been free of controversy. Several Silicon Valley companies have accused DeepSeek of improving its systems through “distillation,” a technique in which a model is trained on the outputs of a more capable one, in this case allegedly American models accessed through their public interfaces. DeepSeek has not publicly addressed those allegations in detail.
The United States has sought to slow China’s progress through export restrictions on advanced chips, while Beijing has poured public money into the sector. Open distribution has become an economic strategy in its own right: cheap, accessible models spread quickly into robotics, logistics and manufacturing, generating data that feeds the next round of improvements.
Soft power through software
The effect is global. A New York Times analysis published this week noted that developers with limited budgets, from Lagos to Kuala Lumpur, are increasingly building on Chinese open models because they are inexpensive and easy to adapt, and estimated that such models accounted for roughly a third of global AI usage over the past year, with DeepSeek and Alibaba’s Qwen family leading.
Analysts describe open source as a form of technological soft power. By making their models the default foundation for developers worldwide, Chinese companies gain influence over how AI is built and deployed far beyond their home market.
