# Qwen Leads in Open-Source: 942 Million Downloads and Dominance in Fine-Tunes
Alibaba's Qwen family of models has reached 942.1 million downloads on Hugging Face by March 2026, more than doubling Llama's 476 million. The ATOM Report, published April 8 on arXiv, analyzes 1,500 key open-source models and documents a market shift toward Chinese developments.
Qwen pulled ahead in downloads in September 2025 with a near tie: 325.4 million versus Llama's 323.7 million. By February 2026, Qwen's monthly downloads hit 153.6 million, establishing it as the de facto standard for developers.
Dominance in Fine-Tunes
Fine-tunes—custom adaptations of base models—offer a clearer picture of real-world production use. Qwen's share of new fine-tunes grew from 1% in January 2024 to 69% in February 2026. Llama fell from 44% (August 2024) to 11%.
This distinguishes casual downloads from actual deployment: fine-tunes require integration into pipelines, testing, and optimization for specific tasks.
- Qwen Growth: +68% market share over 2 years.
- Llama Decline: -33% from peak.
- Other Models: Collectively under 10% of new adaptations.
Comparison with American Models
American players (NVIDIA, Ai2, IBM, and others) racked up 56 million downloads combined against Qwen's 942 million. Chinese models overtook Western counterparts in summer 2025 and keep widening the lead.
Qwen 2.5 (September 2024) delivered a range from 0.5 to 72 billion parameters under Apache 2.0. It rivals proprietary solutions in quality while remaining fully open. DeepSeek in early 2025 reinforced the trend: superior development efficiency with fewer resources.
The open-source AI ecosystem has shifted from Silicon Valley to China—Hangzhou and Shenzhen are the new hubs.
| Metric | Qwen (mil) | Llama (mil) | USA Total (mil) |
|--------------------|------------|-------------|-----------------|
| Downloads (Mar 2026) | 942.1 | 476 | 56 |
| Fine-Tunes (Feb 2026, %) | 69 | 11 | <10 |
Factors Behind Qwen's Success
Qwen 2.5's architecture is optimized for scaling: smaller versions (0.5B) fit edge devices, while larger ones (72B) power high-load inference servers. The Apache 2.0 license enables seamless commercial use without restrictions.
Developers highlight its stability in long-context tasks and multimodal scenarios. Compared to Llama, Qwen delivers a smaller footprint with equivalent perplexity on benchmarks like MMLU and GSM8K.
- Wide range of model sizes.
- Open license without restrictions.
- Competitive quality metrics.
- Active ecosystem of tools on Hugging Face.
Key Takeaways
- Qwen controls 69% of new fine-tunes—a key indicator of production use.
- Chinese models lead the US by 17x in downloads.
- The trend kicked off with Qwen 2.5 and DeepSeek in 2024–2025.
- Downloads signal interest; fine-tunes reflect real deployments.
- The ecosystem is shifting to Asia, reshaping base model choices.
One open question is conversion to enterprise deployments: big companies often favor APIs or private setups, which aren't visible in Hugging Face's public stats. For mid- and senior-level developers, it's a cue to reassess the stack—Qwen cuts vendor lock-in while delivering top performance.
— Editorial Team
No comments yet.