Analysis

Local AI Stands on Chinese Open-Weight Models, and Qwen's Pledge Skips the License

Two thirds of local LLM downloads trace to Alibaba, which uploaded almost none of them. Qwen's pledge names versions, but the license is what to watch.

MesmerTools•November 13, 2026•3 min

On November 2, Qwen's head promised that Qwen 4 and Qwen 5 will stay open. Qwen is Alibaba's model family. Local AI people breathed out. I wanted to know how much rides on that sentence, so I traced the most-downloaded local LLMs of the last 30 days back to whoever made the weights. About two thirds go back to Alibaba.

I believe him. But he promised version numbers, and those downloads stand on a license file he didn't mention. If you run them at home, you're fine. If you ship a product on them, that file is your real dependency.

An industrial tap pouring a glowing amber stream of model weights into a small consumer GPU, with a wax seal hanging from the tap handle and the headline Who holds the tap?

Alibaba uploaded almost none of its two thirds

Of those 66 points, Alibaba's own repos are about 4. The rest come from Unsloth and some 80 other uploaders, and a good chunk of it is abliterated or uncensored builds. None of them had to ask, because most of those weights are Apache 2.0, which lets anyone reuse them. Stare at the first column.

Everything except Alibaba dims, then the LLM column splits by uploader, and that one bright sliver is all it posted itself (one month of downloads, not users).

The license is what holds that pile up. My guess is the pledge is cheap to keep: publishing weights adds little to what Alibaba already spends, and its money comes from cloud. Its June quarter shows which side earns.

Only one of the many streams reaches Alibaba's cloud, yet cloud is the side of the company that turns a profit.

But Alibaba already stopped once, and that wasn't the scary part

It published its video model Wan 2.2 under Apache in July 2025, then shipped 2.5 and 2.6 as APIs only. Sixteen months on, it's still the most-downloaded video model on Hugging Face, and the new work moved to other labs.

Wan keeps the crown on old weights, while Lightricks' LTX and MiniMax's H3 take 90% of downloads among repos created in 2026.

So a stop plays out as a freeze: you keep what you have, and someone fills the gap if it's worth filling. In video, Lightricks was there to fill it. Text has nothing comparable, since Google's Gemma, the biggest non-Chinese family, is under 9% of downloads. I'd bet on another Chinese lab.

So the license on Qwen 4's open-weight model is what I'm watching

A license can turn without anyone breaking a promise. Qwen-Image went from Apache to research-only, the weights are still up, and it's still the most-downloaded image model in my count. Alibaba's text models have split too: the mid-size line has stayed on Apache, while its biggest new models came with strings.

DeepSeek and the mid-size Qwen line stay on Apache or MIT while the lines that earn money pick up chains.

Qwen 4's 27B is expected to be the open tier, but as of November 2026 there are no weights or license to read. If it went the way of Qwen-Image, the community builds would probably survive. Products built on it wouldn't.

At home, change little. Download the Apache builds you rely on now, because that grant is perpetual and a stricter Qwen 4 can't reach back. If you ship a product, build on something already under Apache, like Qwen3.8 27B. What I still don't know is what the license on Qwen 4's 27B will say.

More from the Blog