If you downloaded a powerful AI model and ran it on your own laptop in 2026, there is a good chance it came from China. Alibaba's Qwen, DeepSeek, and Zhipu's GLM are now among the most-downloaded open models on the planet. For a Western developer, that is a quiet shift worth understanding.
Key takeaways
- "Open weights" ≠ "open source." Most Chinese models release trained weights (the learned parameters) with a license, but not always the full training code or data. Know which you are getting.
- China is, by several counts, the largest publisher of open models in the world — a strategic choice, not an accident.
- The two big names to know: Qwen (Alibaba, broad and multilingual) and DeepSeek (efficient, research-driven). See What is DeepSeek? for the deeper story.
- Open release drives adoption and talent, which in turn feeds the lab's ecosystem and future models.
- For developers outside China, these models are often free to use and fine-tune under permissive licenses — a real alternative to closed US APIs.
What "open weights" actually means
A model is a pile of numbers (weights) learned from data. If a lab releases those weights, you can download them and run inference yourself. That is "open weights."
True open source would also mean the training code and the dataset are public, so you could reproduce the model from scratch. Most Chinese releases stop at weights-plus-license — enough to use and adapt, not enough to rebuild. Both are hugely useful; just don't confuse them when you choose a model for a product.
The main players
- Qwen (通义千问) — Alibaba's family, spanning tiny phone-sized models to large ones, with strong multilingual coverage. It is the default "open model" for many global developers.
- DeepSeek — known for squeezing top-tier performance from less compute, and for a research culture that publishes aggressively. Its open-weights move reshaped pricing across the industry; details in DeepSeek's open-weights release.
- GLM (智谱) — Zhipu's line, strong in reasoning and often used in Chinese enterprise.
- Plus a long tail: Baidu, Moonshot (Kimi), and others, each open-releasing at least one tier.
The surge is real and recent — NeuroAI tracked a dramatic acceleration in China's model output over eight weeks.
Why give models away?
It sounds backwards: spend millions training a model, then hand it out. The logic:
- Adoption is the moat. Every app built on your model is a reason to stay in your ecosystem — and to buy your cloud or chips later.
- Talent and feedback. Open release pulls in developers and bug reports you could never hire for.
- Policy alignment. Open models fit national goals of broad AI adoption across industry, described in China's open-source strategy.
So "free" is a business model, not charity.
What it means for you
If you are a developer outside China, these models let you:
- Run capable AI without sending data to a US cloud.
- Fine-tune a base model on your own domain for a fraction of training cost.
- Avoid per-token API bills for many workloads.
The catch is compliance: read each license. Some forbid certain commercial uses or require disclosure; a few carry geopolitical export nuances. None of this is legal advice — just read the license file before you ship.
Honest limitations
This is an explainer based on public model releases and NeuroAI's reporting, not a benchmark shootout. "Largest publisher of open models" reflects counts cited in 2025–2026 industry surveys and can shift quarter to quarter. License terms change; always verify the current license on the model's official repository before commercial use. Performance rankings are contested and task-dependent, and "open weights" coverage varies by model tier — smaller tiers are often more open than the flagship.
Sources
Alibaba Qwen and Zhipu GLM official model cards and licenses; DeepSeek open-weights announcements covered in DeepSeek open weights; NeuroAI reporting on China's eight-week model surge and open-source strategy; industry surveys counting national shares of open-model releases (2025–2026).