A clip went round this week — half a minute of business news, briskly delivered. Alibaba's Qwen models had passed three billion downloads in six months, making Qwen the most popular AI model in the world, ahead of Meta and Google, according to a report from Hugging Face.
Nearly all of that is right. The gaps are worth the five minutes, because one of them is the most common way a business owner gets handed a bad number.
What the report actually says
Hugging Face publishes a twice-yearly survey of open models. Its summer edition counts Qwen downloads for the first seven months of 2026 at 2,045 million.
The three-billion figure is not in it. That number is Alibaba's own, given in a company statement and picked up by the coverage, which then cited the Hugging Face report beside it.
So a company's self-reported figure about itself arrived wearing a third party's credibility. Nobody faked anything. Two true sentences were set next to each other, and the join is where the meaning changed.
It also stops being a fair comparison. Alibaba's number is worldwide and self-reported; the competitor numbers come from one platform's own measurements. Alibaba runs its own model-distribution hub as well, so its global total counts channels that a Hugging Face figure cannot see. The two sides of the comparison were not counting the same thing in the same place.
On the competitors, I will be straight with you: the figures that circulated were 418 million and 227 million for Google and Meta, attributed to this report. I went looking for them in it and could not find them. They may well be right — a direct query of the platform would produce something like them — but a number that everybody sources to a document that does not contain it is the whole subject of this post, so that is all I will say about them.
The uncomfortable part: the conclusion was right anyway
Qwen is the most-downloaded open-model family. The report says so on its own count, and not narrowly. Qwen-based derivatives on the hub number 151,448 — 2.6 times Meta's entire footprint there. Alibaba claims roughly double that figure again, but even the conservative count is a rout.
The underlying shift is real too. Through 2026 the largest open model released each month by a Chinese lab ran from 754B to 2.78 trillion parameters, while American open releases stayed under 130B in five months out of seven.
So the headline was directionally correct and the evidence supports it. Only the margin was inflated.
That is the most dangerous shape a wrong number can take. Nobody audits a figure that confirms something they already believe. A wrong number attached to a wrong conclusion gets caught, because the conclusion looks off and somebody checks. A wrong number attached to a right conclusion survives, gets repeated, and ends up in a deck.
Now the part that affects what you buy
Set the attribution aside and assume every figure was perfect. The ranking still would not tell you what it sounds like it tells you, because of what a download is.
Hugging Face says this itself, in the same report the league table comes from: downloads indicate activity on its own hub and "do not capture API usage, private deployments, or models distributed through other channels." The metric's own publisher tells you not to read it as global usage.
It gets sharper. In that report:
- 83% of all-time downloads go to models under a billion parameters, and 1% to everything above 100B. Restricting to 2026 alone barely moves it.
- 85.6% of models on the hub have fewer than 200 lifetime downloads.
- A small sentence-embedding model — not a chatbot, nothing anyone argues about on the internet — was pulled 1.55 billion times in seven months.
Sit with that last one. A single unglamorous component got download numbers in the same league as the entire Qwen family, because it is a dependency. Automated pipelines re-fetch it on every build. No human chose it that many times.
That is what the metric measures: what scripts pull. Not what a business evaluated, paid for, or put in front of a customer. A download is not a user. It is not even a person.
Does 'most downloaded' mean 'best for my business'?
"Most downloaded" is a popularity contest, and popularity contests are a fine way to find out what is interesting and a bad way to decide what to run. If you are choosing a model for something in your business, the league table cannot answer any of the questions that decide whether it works:
- Does it hold up on your data — your vocabulary, your edge cases, the awkward phrasing your customers actually use?
- Does it hold up at your latency and your cost, at the volume you have, not the volume in a benchmark?
- Does it hold up under your constraints — what you are allowed to send where, what has to be auditable, what happens when it is confidently wrong?
Those are answered by evaluation against your own cases. That is a few days of work, and it is boring, and it is the difference between a feature that works and a demo that worked once.
My own receipts, since this would be cheap otherwise
It would be easy to write this as a lecture. So here is me doing the same thing, this week.
Every number on this site comes out of a register — one file, one row per figure, each with a source. Nothing reaches a post unless it is in there. One of those rows counts the register itself.
That row said 60 when the file held 62. I caught it by luck: a script printed a different total next to it and the two disagreed on screen. Had it happened to match, I would have published a wrong number into the very system whose job is catching wrong numbers. Then, writing this post, the register grew again — it now holds 71 facts — and I had to correct that row twice more in one afternoon.
The lesson is not that I am careless. It is that a number describing something that changes is wrong by default, and the only durable fix is to derive it rather than store it. Mine is still stored. That is a real open problem in my own tooling, and I would rather say so than let the receipts do all the talking.
This is also the honest reason I check figures mechanically instead of carefully. Careful reading is how a boast about projects delivered, and another about client satisfaction, ended up on an earlier version of this site. Both were invented. Both read perfectly. Neither had happened, and I did not catch either by rereading them — I caught them by building something that refuses a figure with no source.
The short version
Qwen leads. The three-billion figure was the company's, not the researchers'. The comparison was not like-for-like. And none of it answers whether any particular model will work for you, because downloads were never a measure of that.
If you have an AI feature in your business and nobody has yet checked how it behaves on your own awkward cases, that is the gap worth closing — not the question of which model is winning.
Run the free website check — a different question, same principle: what is broken on your site that nothing is telling you about?
If you have an AI feature in front of customers, the AI Reliability Audit is $2,500 — your feature reviewed before it ships a wrong or unsafe answer, and you keep the eval harness so you can run it again after every change.
Either way, tell me what you have got. The first conversation is free, and I am one engineer, so you will be talking to the person who does the work.




