State of Open Source on Hugging Face: Spring 2026

The open source AI ecosystem has experienced rapid growth, with the number of users, models, and dataset repositories nearly doubling. As of 2025, Hugging Face has grown to 13 million users, over 2 million public models, and more than 500,000 public datasets, signaling a shift from passive consumption to active participation through the creation of fine-tuned models, adapters, and benchmarks.

Ecosystem Concentration and Corporate Adoption

Despite overall growth, the ecosystem remains highly concentrated: 0.01% of models (the top 200 most downloaded) account for 49.6% of all downloads, while approximately half of all models have fewer than 200 downloads.

Corporate adoption of open source is increasing, with over 30% of Fortune 500 companies now maintaining verified accounts on Hugging Face. Big Tech investment is evident, with NVIDIA emerging as the strongest contributor. The report notes that downstream value created by open artifacts often exceeds production costs, providing organizations with greater flexibility and lower costs compared to exclusively closed systems.

Geographic Shifts and the Rise of Chinese AI

China has surpassed the United States in both monthly and overall model downloads. In the past year, Chinese models accounted for 41% of all downloads. This shift was accelerated by the viral release of DeepSeek's R1 model in January 2025, leading to a surge in open releases from organizations like Baidu (which went from zero releases in 2024 to over 100 in 2025), ByteDance, and Tencent.

While the U.S. and Western Europe continue to contribute via large labs like Meta and Google, the trajectory of repository growth among Chinese organizations is significantly steeper. Additionally, individual and unaffiliated developers have seen their share of downloads rise from 17% to 39% since 2022, often focusing on quantization and adaptation of base models.

National Sovereignty and AI Initiatives

Open weight models are increasingly viewed as tools for national sovereignty, allowing governments to deploy AI on domestic hardware and fine-tune systems on local data.

• South Korea: Launched the National Sovereign AI Initiative in mid-2025, naming national champions including LG AI Research, SK Telecom, Naver Cloud, NC AI, and Upstage.
• Europe: Switzerland's Swiss AI initiative and various EU-funded projects prioritize similar goals, while the UK follows the "public money, public code" principle.

Model Popularity and Technical Trends

Shift in Popularity

Community attention has shifted from a predominantly U.S.-developed landscape (Meta's Llama family) to an international mix, with China's DeepSeek-R1 currently at the top of the most-liked models.

Derivative Models

The Qwen family from Alibaba has seen massive adoption, with over 113,000 derivative models (and over 200,000 when including all models tagging Qwen), surpassing the combined derivatives of Google and Meta.

Accessibility and Model Size

Smaller models are downloaded and deployed at significantly higher rates than large systems due to constraints in cost and latency. While the mean size of downloaded models rose from 827M parameters in 2023 to 20.8B in 2025 (driven by MoE and quantization), the median size increased only marginally from 326M to 406M parameters.

Emerging Sub-Communities

Robotics

Robotics is the fastest-growing sub-community on Hugging Face. Robotics datasets grew from 1,145 in 2024 to 26,991 in 2025, becoming the largest dataset category on the Hub, surpassing text generation. Key developments include the LeRobot library and the Learning to Drive (L2D) multimodal dataset.

AI for Science

Open models are increasingly applied to protein folding, molecular dynamics, and drug discovery. While frontier AI companies have dedicated science teams, much of the current community focus is on literature discovery and large-scale interdisciplinary coordination.

Compute and Hardware Optimization

Development is shifting toward cross-hardware compatibility. While NVIDIA GPUs remain dominant, support for AMD hardware is expanding, and Hugging Face launched the Kernel Hub in 2025 to optimize kernels for both NVIDIA and AMD. Simultaneously, Chinese models are being released with explicit support for domestic chip architectures to reduce reliance on foreign cloud infrastructure.