China's massive open models outshine the US's smaller ones

Summary

Hugging Face's 2026 summer report documents that Chinese labs have consistently outpaced their US counterparts in frontier open-model releases. While Chinese teams offered models ranging from 754 billion to 2.78 trillion parameters, the vast majority of US teams stayed below 130 billion parameters. The report also shows that the bulk of actual download volume is driven not by hyped frontier models, but by small, stable models. The data further reveals a growing divergence in licensing. Recent large models like Kimi K3 and Qwen 3.8 2.4T introduced non-commercial restrictions, contrasting with the Apache 2.0 and MIT licenses that dominated the previous year.

Why it matters

The report quantifies a structural shift. Open-weight AI is no longer solely a US-led endeavor. Economic value is moving from positions in frontier benchmarks to the small-model infrastructure layer that pipelines actually rely on. A critical finding for developers and policymakers is that just 1.5% of repositories account for 99.2% of downloads. This highlights how concentrated and fragile the real usage base of the open ecosystem is, especially as the number of open models approaches 3 million.

Highlights

• Chinese frontier labs (Moonshot, MiniMax, Xiaomi, Z.ai) release almost no models under 70 billion parameters. This results in most initial-contact models being too large for local execution for the majority of developers.
• Only one model appears on both the top 25 most downloaded and top 25 most liked lists. Thirteen of the top 25 most downloaded models are from 2022. No 2026 model appears on this list.
• Qwen's full-spectrum strategy, spanning 1 billion to 2.4 trillion parameters, generated approximately 2.045 billion downloads. Moonshot's frontier-only portfolio received 37 million downloads. There is a roughly 55-fold difference between these two figures.

Critical take

The report is inherently self-referential by design. Since Hugging Face audits its own hub, metrics like 'downloads' and 'likes' reflect platform-specific behaviors rather than true ecosystem adoption. The authors explicitly acknowledge that they previously conflated attention with usage. The licensing section cuts off mid-sentence. This leaves the report's most significant claim—that a shift toward non-commercial use and revenue sharing is emerging in the largest models, contradicting the permissive licensing trend—underdeveloped and difficult to weigh.