2.26 Million AI Models and Counting — China Overtakes the US, and Solo Developers Surpass Big Tech
Hugging Face's Spring 2026 State of Open Source AI report reveals that China has overtaken the US with a 41% download share, and independent developers have surpassed big tech companies for the first time. Robotics datasets exploded 24x in just one year.
The world's largest open-source AI platform, Hugging Face — where anyone can freely upload and download AI models — has published its Spring 2026 State of Open Source Report. The bottom line: AI is no longer just for big corporations. Individual developers and China have caught up with US Big Tech, and robotics is showing explosive, unexpected growth.
• 2,264,880 AI models — 22x growth from 100,000 just three years ago
• 11 million users and 500,000 datasets surpassed
• China's download share hits 41%, overtaking the US
• Individual developers' share reaches 39%, surpassing big tech (37%) for the first time
• Robotics datasets grew 24x in one year (1,145 → 26,991)
From 100K to 2.26M — An AI Model Ecosystem That Grew 22x in Three Years
In 2022, there were around 100,000 AI models on Hugging Face. That figure crossed 500,000 in 2023, when Meta released LLaMA 2 (a powerful open-source language model), and surpassed 1 million in 2024, driven by Alibaba's Qwen2 and the image-generation model Flux 1.0. Then in 2025, with the arrival of DeepSeek-R1, the count reached 2.26 million.
But the raw numbers can be misleading. The top 0.01% (roughly 200 models) account for 49.6% of all downloads, while more than half of all models never reach 200 downloads. It's the same dynamic as app stores, where a handful of apps capture the vast majority of downloads.
China Overtakes the US — 41% of All AI Downloads
The most striking shift is China's rise. Until 2024, the US was the undisputed leader in downloads. In 2025, China surged to 41% of all downloads, overtaking the US. The catalyst is clear — the public release of DeepSeek-R1 in January 2025 ignited the entire Chinese AI ecosystem.
Chinese companies' participation in open-source AI also exploded.
• Baidu: 0 → 100+
• ByteDance (parent company of TikTok): 8–9x increase
• Tencent (parent company of WeChat): 8–9x increase
• MiniMax: Shifted from closed-source strategy to open-source
Alibaba's Qwen Has More Derivative Models Than Google and Meta Combined
Derivative models — models that other developers create by modifying an existing base model for their own purposes — are a key indicator of an AI's real-world influence. Alibaba's Qwen series has generated 115,000 derivative models. That's more than Google (72,000) and Meta (46,000) combined.
This means Qwen isn't just a "popular model" — it has become foundational technology that other developers adapt for their own specific needs. In Korea, for instance, Qwen-based models are increasingly the go-to starting point for building Korean-language AI applications.
A Historic Moment: Individual Developers Surpass Big Tech
In 2022, big tech companies accounted for roughly 70% of AI model downloads, while individual developers held just 17%. By 2025, individual developers' share had climbed to 39%, surpassing big tech (37%) for the first time ever.
What does this mean? Creating and deploying AI no longer requires massive capital or thousands of GPUs (graphics processing units — the specialized hardware used to train and run AI models). A virtuous cycle is now fully underway: individuals download open-source models like Qwen or Llama, fine-tune them (adjust them for a specific task or domain), and re-upload their results for others to use.
An AI Model's Shelf Life Is About 6 Weeks
Here's an intriguing finding. The window during which an AI model captures meaningful attention — its effective "shelf life" — is about six weeks on average. When a new model launches, it generates a burst of excitement, but is quickly forgotten once a better successor arrives.
There are exceptions. DeepSeek managed to sustain attention by releasing rapid successive updates — V3, then R1, then V3.2. The report calls this a "continuous update strategy" and notes that, just like apps, consistent version management can make or break an AI model's staying power.
The AI Models People Download Are Now 25x Larger Than Two Years Ago
In 2023, the average size of AI models people actually downloaded was 827 million parameters (a parameter is like a connection in an AI's "brain" — more parameters generally means a more capable model). By 2025, that figure had grown to 20.8 billion — a 25x increase.
However, the median size (the most commonly used model size) only grew modestly, from 326 million to 406 million parameters. This tells us that most practitioners still prefer small, fast models, but a subset of power users downloading very large models is pulling the average up significantly.
Robotics Datasets Have Become the #1 Category on Hugging Face
The most unexpected trend is in robotics. Training datasets for robot AI jumped from 1,145 (ranked 44th overall) in 2024 to 26,991 (ranked 1st overall) in 2025 — a 24x increase. That's more than five times the volume of the second-place category, text generation (roughly 3,000 datasets).
This aligns with the wave of robot-specific AI chips and platforms unveiled at NVIDIA GTC 2026. The data confirms what many have been predicting: AI's next frontier isn't a screen — it's the physical world. Manufacturing, logistics, and healthcare are likely to feel the impact first.
Korea Steps Up — A National AI Initiative and Five Champions
The report highlights Korea's "National AI Sovereignty Initiative" as a notable case study. Launched in mid-2025, the program selected LG AI Research, SK Telecom, Naver Cloud, NC AI, and Upstage as national AI champions.
The results are already showing: in February 2026, three SK AI models simultaneously appeared on Hugging Face's trending list. Korea is beginning to establish itself as a genuine exporter of AI models on the global stage.
30% of Fortune 500 Companies Are Already on Hugging Face
Enterprise adoption is accelerating. More than 30% of Fortune 500 companies now have an official presence on Hugging Face, and startups are increasingly treating open-source models as off-the-shelf components rather than paying for proprietary AI services.
The report estimates that open-source AI models can reduce costs by 10x to as much as 1,000x compared to commercial alternatives. A startup previously spending hundreds of dollars a month on the ChatGPT API could see costs drop dramatically by switching to a self-hosted model built on Qwen or Llama.
3 Takeaways From This Report
Rather than simply using AI built by a big tech company, downloading an open-source model and customizing it for your own work is becoming a genuinely practical option. If you have some coding experience, you can head over to Hugging Face and try it yourself today.
Chinese open-source models like Qwen and DeepSeek now match US models in both capability and ecosystem scale. For Korean-language AI projects in particular, Qwen-based models are often the most practical and effective starting point.
Beyond the era of text chatbots, we are entering the age of "physical AI" — where AI perceives and manipulates real-world objects. The data shows the groundwork is already being laid at scale. Manufacturing, logistics, and healthcare will be the first sectors to feel the transformation.
The full report is available on the official Hugging Face blog, complete with charts and interactive data that make it easy to grasp where the AI ecosystem is heading.
Related Content — Get Started with AI | Free Learning Guide | More AI News
Stay updated on AI news
Simple explanations of the latest AI developments