Chinese AI Models Close Gap with US Frontier Systems, Study Shows
A new study by the UK's AI Security Institute indicates that China's leading open-weight AI models are catching up with American frontier systems at an accelerating pace. Frontier models are the most advanced AI systems currently available, while open-weight models allow developers to access and customise the underlying weights, enabling independent operation rather than relying on the company's proprietary service.
According to the research, Chinese models such as Z.ai's GLM-5.2 and DeepSeek V4-Pro now perform at a level comparable to frontier closed models released just four to seven months earlier. For most of 2025, the gap was estimated at six to ten months. This trend reflects a broader pattern that AI experts have observed over recent months.
"The open-source gap is collapsing exponentially. What took 10 months now takes four, and soon there won't be a gap at all," said Srinivas Padmanabhuni, CTO of AiEnsured. Sagar Vishnoi, co-founder of Future Shift Labs, echoed this view: "The gap between frontier closed models and leading open-weight models is shrinking, indicating that breakthroughs are diffusing far more quickly across the global AI ecosystem than they did even a year ago." According to Vishnoi, this is likely to intensify competition, lower barriers to entry for developers, and accelerate innovation beyond a handful of frontier AI companies.
This week, another Chinese AI model, Kimi K3, developed by Beijing-based startup Moonshot AI, reached the top of Arena's AI coding leaderboard within hours of its launch, becoming the first Chinese model to claim the No. 1 spot. The leaderboard is widely followed by AI researchers and developers as a measure of coding performance. The achievement has reignited debate over whether the US can maintain its lead in AI. For companies like OpenAI and Anthropic, the concern is not just that Chinese models are improving, but that many are significantly cheaper and open-weight, allowing developers to download, customise, and run them on their own infrastructure instead of paying for proprietary cloud-based services. This combination could make them increasingly attractive to enterprises seeking to reduce AI costs.
The comparison many in Silicon Valley are making is with DeepSeek, which earlier this year released a highly capable model at a fraction of the cost of American rivals, briefly wiping hundreds of billions of dollars off the value of major US technology companies. Some believe Kimi K3 could have a similar impact. Anastasios Angelopoulos, who runs the Arena benchmark, said the model could force investors to rethink assumptions if businesses increasingly choose free, customisable Chinese models over expensive proprietary ones. David Sacks, the White House's AI adviser, argued that the latest developments show the US cannot afford to slow down AI development with excessive regulation. Others, like Dean Ball, who recently served as an AI adviser in the White House before joining OpenAI, described Kimi as a genuinely capable model, evidence that Chinese AI companies are no longer simply copying American systems.
The excitement around Kimi K3 has been so intense that Moonshot AI temporarily paused new subscriptions less than two days after launch, as demand pushed its computing infrastructure close to capacity. The company said existing paid users would continue to receive uninterrupted access while it adds more GPUs.
