Western open-weight models still can’t beat China

New open-weight models from American and European companies aim to win over users of Chinese models, yet they still trail the leading Chinese systems on performance despite having greater resources.

Last week, the American startup Reflection released its first open-weight model, and European labs Aleph Alpha and Mistral put out new ones too, all targeting users who might be worried about security with Chinese models. But experts at an event in San Francisco said these new models fall short of the dominant Chinese open models on performance, even with more resources.

Affordability and fit decide the market

Daniel Björkegren, an assistant professor at Columbia University, pointed out that for businesses and governments in lower-income countries, cost comes first, along with the expertise to customise models for education, healthcare and farming. American models are tuned for American workloads and often do not work well in remote areas.

On the OpenRouter leaderboard, four of the top five models last week were Chinese: DeepSeek, Z.ai, Xiaomi and Tencent. On LiveBench, DeepSeek made the top five for the first time, and the best Chinese models lagged behind U.S. rivals by only 3% on benchmark scores.

Open models as geopolitical leverage

China's leader has called open models key to reaching global AI leadership by 2030. This year China set up the World Artificial Intelligence Cooperation Organization to promote access to Chinese models and training in developing nations.

Research fellow Kristy Loke said China has talked about open innovation for a long time but faced pushback, and open models let people download, use and own them, which breaks down the trust problem. It is good for soft power and domestically too. The U.S. directive suspending access to some Anthropic models for foreign nationals this year reinforced to governments that they should not depend entirely on American systems.

The U.S. has resources but won't release frontier models

Graham Webster, a research scholar at Stanford, said open models from Meta and Nvidia do not come close in capability. If you need relatively frontier capability right now, that is Chinese open models only. American companies simply do not release open models at that performance level, so we cannot see if users would prefer a U.S.-made one.

Chip restrictions have forced Chinese companies to build leaner, more efficient models. They still face compute limits and uncertain monetisation, but being able to do some tasks 50% to 90% cheaper than leading American frontier models is a strong draw.