According to a 2025 study from AI model marketplace OpenRouter, Chinese open-source AI models now account for nearly one-third of global AI use, a market situation that would have been unimaginable just two years ago. DeepSeek is the leader in that segment, followed by Alibaba’s Qian family. The numbers tell a story that Silicon Valley is no longer in control.
DeepSeek’s newly released v4 model outperforms every other open-source system in generating computer code, according to independent testing by Walls AI, which benchmarks AI performance across the industry. This puts it ahead of fellow Chinese startup Moonshot AI’s recently released KM2.6, although Walls AI CEO Ryan Krishnan says the two are “basically head-to-head” across a wide range of functions.
The difference that matters most is still the difference between China and the US. V4 is ahead of Anthropic and OpenAI’s top proprietary coding models, but the distance is closing with each release cycle. that trajectory is exactly the same
The timing of DeepSeek’s V4 launch is not coincidental. In the lead up to the release, OpenAI and Anthropic publicly accused DeepSeek of using distillation, a technique where engineers interrogate rival models millions of times to replicate their behavior without building them from scratch. Both companies defined it as inappropriate appropriation of their proprietary technology.
DeepSeek has not formally responded to those allegations. But the allegations reflect a broader concern: Open-source Chinese models are advancing not only through independent research, but also by learning directly from the closed systems they are racing to defeat.
China’s adoption of open-source AI is not accidental generosity; This is a well thought out strategy. A 2025 study by an advisory body to the US Congress found that freely available Chinese models have spread rapidly across domestic industries including robotics, logistics and manufacturing, generating real-world data that helps further improve systems.
Globally, developers from Lagos to Kuala Lumpur are adopting Chinese models because they are cheaper to run. Malaysia’s deputy communications minister said last year that the country’s sovereign AI infrastructure would be built on DeepSeek’s technology, a solid sign of the geopolitical base built by open-source.
Three consecutive US administrations have imposed export controls on advanced chips to limit Chinese AI development. Despite this, Alibaba’s Quen model family has surpassed one billion downloads. ByteDance plans to invest $11 billion in AI infrastructure in 2024 alone. Chip restrictions have slowed down some capabilities but not broken the speed.
Kevin Xu, founder of Interconnected Capital, a hedge fund focused on AI investments, puts it bluntly: “Open source is the soft power of future technology.”
