Chinese open models close the capability gap while winning developers through distribution

The Information’s Eric Bellomo reports that Chinese open models now reach about 95% of the performance of leading closed models in the publication’s analysis, while Chinese labs increasingly drive both capability convergence and developer adoption.

The number needs careful interpretation: an aggregate benchmark gap does not mean identical performance on every task, equal reliability in tool use or comparable safety. “Open source” in model markets also often means downloadable weights under varying licences, not full disclosure of training data and methods. Even so, a small enough quality gap changes purchasing decisions because developers gain lower serving costs, local deployment, customization and freedom from a single API provider.

The adoption evidence is broader than one score. Brookings notes that Chinese models have overtaken US models in cumulative Hugging Face downloads; derivative uploads built on Chinese foundations have also moved ahead, and Alibaba’s Qwen has displaced Meta’s Llama as the leading open family by popularity. US chip controls helped push Chinese labs toward efficiency, while open distribution converts that constraint into reach.

The AI contest is not only about who owns the best frontier benchmark. If developers standardize on Chinese model families because they are cheap, modifiable and deployable anywhere, those ecosystems can shape tools and infrastructure even while US labs retain a narrow capability lead.