THE CRUNCH
A wave of new open-weight AI models from Western companies, including Reflection's first release and fresh models from Mistral and Aleph Alpha, is aimed at prising away users of dominant Chinese open models who worry about security. Yet experts speaking at a Rest of World event in San Francisco said the newcomers still fall short of the leading Chinese systems on performance, despite their developers having greater resources. Open-weight models, which can be downloaded, customised and run by their users, offer more control than closed offerings from OpenAI, Anthropic and Google at significantly lower cost. That combination has made Chinese open models the default for many businesses and governments, particularly in lower-income countries.
The numbers back up the gap. The OpenRouter leaderboard of large language models had four Chinese models, from DeepSeek, Z.ai, Xiaomi and Tencent, in its top five last week. LiveBench, which ranks models on various benchmarks, recently placed DeepSeek in its top five for the first time, with the best Chinese models trailing their U.S. rivals by just 3% on benchmark scores.
Cost and fit matter as much as raw capability. Daniel Björkegren, an assistant professor at Columbia University, said cost is the primary factor for businesses and governments in lower-income countries, alongside the expertise needed to tailor models for local uses such as education, healthcare and farming. American models, he noted, are tuned for American workloads and often do not work so well when deployed in remote parts of the world.
There is also a trust calculation. Kristy Loke, a research fellow at the AI safety programme MATS, argued that open models help China sidestep the wariness that met companies like Huawei, because users can download and own the models privately, easing concerns over the Chinese state. She called that good for soft power. Meanwhile, a U.S. government directive earlier this year telling Anthropic to suspend access to its Fable 5 and Mythos models for all foreign nationals on national security grounds reminded several governments not to depend entirely on American systems, Björkegren said.
China's ambitions are explicit. Xi Jinping has described open models as key to the country's goal of global AI leadership by 2030, and this year it established the World Artificial Intelligence Cooperation Organization to promote AI adoption and shape governance in developing nations, including access to Chinese models and training for students and workers.