On August 14, 2026, the Qwen models, from China's Alibaba, surpassed 3 billion cumulative downloads on Hugging Face — the repository where the world downloads artificial intelligence models. In the same survey, Google appears with 418 million and Meta with 227 million.
The most-used open model family on the planet isn't American. And the detail that matters to decision-makers isn't nationality: it's the file size.
17 GB on an old graphics card
The Qwen3.8-27B, one of the family's recent releases, surpassed 3 million downloads in its first weekend. It's a file of about 17 GB and runs on already-old graphics cards, like an RTX 3090, or on a laptop with enough memory.
This changes the conversation about AI in the enterprise more than any benchmark. We're not talking about a dedicated server or a monthly contract with a vendor: we're talking about a machine that many companies already have sitting idle in the office.
Why this doesn't contradict what we said before
A few days ago we showed that running a frontier open model in-house is expensive — that case involved a 2.88-trillion-parameter model, which required about 6,060 GB of video memory and a small data center to run.
That still holds true. What the Qwen case shows is that "open model" isn't a single thing. There are at least two categories, separated by a gulf in cost:
- The frontier giants, which compete with the top of the market and require data center infrastructure. Downloading is free; running is not.
- The mid-sized models, in the tens-of-billions-of-parameters range, which fit on a single graphics card and handle specific tasks very well.
The useful question, then, is no longer "open or closed." It has become: what size of model do I actually need for the task I have?
What a mid-sized model running on your own machine is good for
Much of a company's real work doesn't require the smartest model in the world. It requires a model that's consistent, cheap, and doesn't send data outside:
- Classifying and routing. Reading what comes in — email, form, order, complaint — and saying where it goes.
- Extracting information from documents. Pulling structured data from invoices, contracts, reports, forms. Repetitive task, high volume, objective answer.
- Summarizing and comparing. Condensing support history, comparing versions of a document, pointing out what changed.
- Answering based on your own material. Querying your own knowledge base and answering from it.
In these four cases, a mid-sized model running locally is usually enough — with two advantages a frontier model doesn't offer: near-zero cost per query after the hardware is paid for, and data that never leaves your network, which matters for anyone handling sensitive information.
What to check before getting excited
- Downloaded isn't the same as good for you. Three billion downloads measure adoption, not fit for your task. Test it with your own material before deciding.
- The lead comes with a caveat. Specialized reporting points out that the advantage over competitors is smaller than advertised, depending on how downloads are counted. The number is impressive regardless, but it isn't an official scoreboard.
- Someone needs to maintain it. A local model means a machine, updates, and an owner. Without that, you've traded a subscription for a problem.
- Licensing is a detail that bites. "Open" varies case by case; check what the license actually allows before putting it into a product you sell.
The honest summary
The most relevant fact in this episode isn't geopolitical, it's economic: Alibaba has released more than 460 models in the family, which have already spawned around 300,000 variations created by third parties. That means there's now a huge shelf of specialized, free models, and most of them fit on common hardware.
For a small or medium-sized business in Brazil, that matters more than any dispute between labs: the entry barrier for using AI with your own data has dropped to the price of a used graphics card.
Facts verified from the Hugging Face report of August 14, 2026, and coverage by Bloomberg, Fortune, and Business Standard. The comparison of downloads between vendors has a methodology disputed by parts of the trade press.
Perguntas frequentes
What is the most downloaded open AI model in the world?
The Qwen family from China's Alibaba, which passed 3 billion cumulative downloads on Hugging Face on 14 August 2026, ahead of Google (418 million) and Meta (227 million). One caveat: part of the specialised press argues the lead is smaller than advertised, depending on how each download is counted.
Is it true these models can run on an ordinary graphics card?
Some can. Qwen3.8-27B is a file of roughly 17 GB and runs on older cards such as an RTX 3090, or on a laptop with enough memory. It passed 3 million downloads on its opening weekend. This does not apply to frontier open models, which require data-centre infrastructure.
So is running AI in-house cheap or expensive?
It depends entirely on model size, and that is where nearly all the confusion starts. A frontier open model with trillions of parameters can demand thousands of gigabytes of video memory and a small data centre. A mid-sized model, in the tens of billions of parameters, fits on a single graphics card. The useful question is not open versus closed, but what size of model your task actually needs.
What is a mid-sized model running locally good for?
High-volume repetitive work: classifying and routing what comes in, extracting structured data from documents, summarising and comparing texts, and answering questions from your own knowledge base. In those cases it usually suffices, with two advantages: cost per query near zero after the hardware, and data that never leaves your network.
What are the risks of adopting an open model in a company?
Adoption is not fitness for purpose: 3 billion downloads say nothing about whether the model suits your task, so test it on your own material. Someone has to maintain the machine, the updates and the operation, or you have traded a subscription for a problem. And check the licence, because open varies case by case and not every licence allows use in a product you sell.
What does this change for small and mid-sized companies?
The entry floor. Alibaba has released over 460 models in the family, which have already spawned some 300,000 third-party variations — meaning there is now a vast shelf of specialised, free models, most of which fit on ordinary hardware. Using AI with your own data no longer requires a data-centre budget.