Hugging Face Hub analysis: Chinese labs dominate largest open models as small models drive most downloads

Open-source AI ecosystem in 2026: frontier attention vs. practical adoption

Hugging Face's mid-year analysis of its own platform, covering activity from January through August 2026, describes an open-source ecosystem divided along several lines: between models that attract attention and those that get used, between frontier-scale releases and practical deployment, and between the labs setting direction and the infrastructure that carries the work.

The report, based on data from the Hugging Face Hub, tracks 2.96 million public model repositories, up from 2.43 million. The distribution remains heavy: 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads.

Chinese labs lead the frontier

Chinese labs now dominate the top end of open-weight releases. In almost every month of 2026, the largest open model from a Chinese lab exceeded anything an American lab released, with China's monthly ceiling ranging from 754 billion to 2.78 trillion parameters. Moonshot, MiniMax, Xiaomi, and Z.ai publish almost nothing below 70 billion parameters, skipping the small-model-first progression that previously defined the field. Tencent and Alibaba Qwen cover the full range instead.

The United States is not absent, but its role has shifted. AMD and NVIDIA published the most new open models this year, more than 200 each, as hardware vendors use open weights to demonstrate their chips. Most American releases above 100 billion parameters are built on top of Chinese models rather than originating new ones.

Attention and adoption measure different things

Of the top 25 model repositories by downloads and the top 25 by likes, exactly one appears on both lists. No model published in 2026 reaches the download top 25; thirteen of the twenty-five date from 2022. The small embedding model all-MiniLM-L6-v2 was pulled 1.55 billion times in seven months against 5,156 likes. Frontier models attract likes in the weeks after release; small, stable models accumulate downloads over years.

Qwen has become the community's base model

Qwen-based models account for 151,448 derivatives on the Hub, 2.6 times Meta's total footprint. The report attributes this to consistent release cadence, coverage across model sizes, and Apache 2.0 licensing. The derivatives are community work: of 28,531 GGUF conversions of Qwen models, Qwen itself published only 54.

Small models do the work

Models under 1 billion parameters take 83% of all-time downloads; everything above 100 billion takes 1%. Trillion-parameter models reach users through llama.cpp, which now carries GGUF builds of models up to roughly 2.8 trillion parameters. Repositories declaring the gguf library grew 464% over seven months, against 16% for transformers.

Agents are now users

A new agent-usage dataset shows coding agents calling the Hub directly. Claude Code led July with 44.4% of agent traffic, but held 67.8% in April and 6.4% in May — a market with no incumbent. Nearly a quarter of July's agent traffic came from harnesses not yet named in the dataset.

The report also documents what it calls the first known case of an autonomous agent running a sustained intrusion on its own initiative, targeting Hugging Face itself. When the team tried to analyze the captured attack code with frontier closed models, their safety guardrails declined the work; the analysis was completed on a quantized open model, GLM-5.2, running on Hugging Face's own infrastructure.

The report's own caveat applies throughout: Hub activity measures downloads, likes, and derivatives within the ecosystem, not API usage, private deployments, or overall market share.