The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+

Source: HuggingFace Date: 2026-02-03 URL: https://huggingface.co/blog/huggingface/one-year-since-the-deepseek-moment-blog-3

Summary

Research/analysis article: HF’s third installment in a series examining China’s open-source AI ecosystem, one year after DeepSeek R1. Key quantified facts: Qwen became the most-derived base model on HF by mid-2025 (113k+ derived models, 200k+ repositories tagging Qwen) — substantially exceeding Meta Llama (~27k) and DeepSeek (~6k). DeepSeek is the most-followed HF organization; Qwen is 4th. Alibaba has nearly as many model derivatives as Google and Meta combined. Most popular HF papers now from Chinese organizations (ByteDance, DeepSeek, Tencent, Qwen).

Implications

Open-weights ecosystem health. Qwen at 113k derived models vs Llama at 27k is the most concrete data point in this post — it means Alibaba’s open-weights model family has become the preferred base for the fine-tuning and adaptation community, surpassing Meta’s Llama which previously dominated derivatives. This is a structural shift in who the open-weights ecosystem builds on.

Model release cadence. Chinese organizations dominating both the derivative model count and the most-read papers on HF represents a sustained reversal of the prior US-centric open-weights landscape. The strategic logic — open-source as deployment and integration strategy rather than capability gating — is a different theory of value capture than US labs have generally pursued, and it’s showing ecosystem results.