Hugging Face report: Chinese open-source models lead in scale, AMD and NVIDIA top US open-source publishers

Hugging Face, the world's largest open-source AI platform, has published a new observation report on the global open-source AI ecosystem, drawing attention to the rising weight of Chinese open-source models and a shift in who drives open releases. The report's core finding: the center of gravity in open-source AI is moving.

According to the report, in most months of 2026 the largest open-source models released by Chinese AI companies have significantly surpassed their US counterparts in parameter count. China's monthly parameter ceiling ranged from 754 billion to 2.78 trillion, while US models stayed below 130 billion in most months over the same period.

Chinese models also stand out in openness: among models with more than 20 billion parameters, over 80 percent use permissive licenses that allow secondary development and commercial deployment. Hugging Face's data suggests Chinese AI firms are attracting developers with open weights and monetizing through API services, cloud computing, hardware adaptation, and ecosystem influence rather than license fees.

Qwen in particular has become a cornerstone of the Hugging Face ecosystem, with more than 150,000 derivative models built on it, far outpacing competitors such as Google and Meta.

The report also found that AMD and NVIDIA, not Google or Meta, are now the two US companies releasing the most new open-source models this year, a sign that the main driving force in open-source AI is shifting from model labs to hardware makers.

For chip vendors, open models have become a lever for hardware sales: releasing models, optimization tools, and hardware adaptation kits expands their chip ecosystems. NVIDIA's Nemotron series showcases GPU synergy through open models, while AMD focuses more on model conversion and performance tuning. Meanwhile, Chinese open-source models are increasingly optimized for domestic chips.

On the application side, small models remain the workhorses, with models under 1 billion parameters accounting for the vast majority of downloads. Hugging Face also believes AI agents are becoming new users of model calls, and a large share of future inference traffic may be executed by agents.

Coming from the platform at the center of the open-source ecosystem, the report's data points to a reshuffling of power: Chinese labs lead on scale and openness, while chip makers have displaced model labs as the top US open-source publishers. Watch whether Qwen's derivative ecosystem keeps growing and whether agent-driven calls reshape how models are distributed.