The Chinese Labs
DeepSeek looked like a surprise, but it was one move inside a much larger one. Chinese labs had been releasing weights steadily for years, and by July 2025 Chinese state media counted 1,509 of the world’s roughly 3,755 public models as Chinese — more than any other country. That count is from mid-2025 and the cadence has not slowed since. The strategy is deliberate, and it is not only about generosity.
The families
Qwen (Alibaba). The broadest line by far: dense and mixture-of-experts models, sizes from tiny to huge, and text, vision, audio, and code variants under one naming scheme. It has become the default base for fine-tuning worldwide, and most newer open models are measured against it. The line has kept moving: the 3.5 generation was followed by 3.6 and then Qwen3.8, and in 2026 Alibaba opened the weights of a Qwen-Max-class model — 2.4 trillion parameters — for the first time. That is a different order of release from putting a mid-size model out for free.
GLM (Zhipu AI, also written Z.ai). Strong general models with an emphasis on agents and tool use, released on a steady cadence.
Kimi (Moonshot AI). Known first for very long context, then for Kimi K2, a large mixture-of-experts model released with open weights and tuned for agentic work.
MiniMax. Efficient mixture-of-experts models with a focus on long context, plus audio and video lines. InternLM (Shanghai AI Lab), Hunyuan (Tencent), Doubao / Seed (ByteDance), and Ernie (Baidu, which had pursued a closed strategy before pivoting) round out the list. Several of them now open their weights as a matter of course.
Why give it away?
Four reasons, none of them charity
Distribution. A free download is the marketing: it earns users, developers, and citations. Ecosystem. Whoever’s model everyone fine-tunes on sets the standard others must follow. Talent. Strong open work attracts researchers. The squeeze. American export controls limited access to the best chips, so efficiency and openness became a way to compete without the biggest clusters — an idea we pick up in the next chapters.
What to watch
A model family is not one model but a lineage — sizes, versions, and modalities that share training choices and a tokenizer. When a family ships under a permissive licence, its tokenizer and chat format quietly become a standard, and everything downstream inherits both its strengths and its quirks. The next chapter widens the view beyond China, and then we look at the constraint that shaped all of it: compute.