Why this exists
English coverage of Chinese AI has a shape to it. DeepSeek, Qwen, Kimi and a handful of other large labs get written up within hours. Below that tier, most of the ecosystem is invisible in English: a robotics company commissioning its arms on a car assembly line, a two-person team open-sourcing a model, a seed round in a field that does not have a name yet. These get covered in Chinese, and then nowhere else.
China AI Compass is a record of that second group.
What is here, and what is not
Every entry is something a company, lab or researcher announced: a product launch, a funding round, an open-source release, a technical result, a partnership. We write down what happened, who was involved, and where it was reported.
We do not publish policy analysis. No readings of regulatory intent, no geopolitics, no commentary on government decisions. Other publications do that work, and some do it very well. This site is a record of what companies are actually shipping, and it is more useful if it stays that.
How events get selected
Two to three hundred Chinese articles come in each day from ten publications: QbitAI, Jiqizhixin, AI Era, 36Kr, Leiphone, TMTPost, Zhidx, GeekPark, ITHome and OSCHINA.
Most are discarded immediately. What survives has to be a specific event rather than a think piece, and it has to have a Chinese company, lab or developer as its subject. Stories about the largest labs mostly get set aside, on the grounds that they are already covered in English within the day. One or two of the biggest stories each day are kept anyway, because a record of Chinese AI that left them out would be a strange record.
What is left goes through the test the site is built around. We take the subject's English and pinyin names and run them against Google News and Bing as an outside check. If the story already has English pickup, it does not belong here and is dropped. If it has little or none, it goes to an editorial screen that drops routine version bumps, marketing and announcements with nothing in them.
That screen is now run by a language model against rules a person wrote, not by a person reading every item. Items it lets through can still be pulled by hand.
On names
A Chinese AI company usually answers to three names: a Chinese one, an official English one, and a pinyin transliteration that appears when nobody checked. The official name is rarely the obvious guess. 智象未来 is HiDream.ai, not "ZhiXiang Future". 它石智航 is TARS. 自变量 is X Square Robot. 启元机器人 is PrimeBOT.
Getting this wrong makes a company unfindable, so each profile lists every variant we know of, and a search for any of them should land somewhere useful. Names are not individually checked against company sites before publication, so treat an unfamiliar English name with some caution.
How these pages are written
Plainly: the English summaries are written by a language model reading the original Chinese article, and the choice of what to publish is automated too. Nobody reads each entry before it goes up.
We would rather say this than have you wonder. Every page links to the Chinese article it came from, which is there so you can check the claim rather than take our word for it.
Corrections
Some of this will be wrong. Amounts get garbled in translation, companies rename themselves, a first-of-its-kind claim turns out to be a third. Where we find an error, the page is corrected rather than quietly deleted, and the original sources stay linked on every entry so that anything here can be checked against them.