The day the Qwen team walked out

What happened on March 4

At about ten past one in the morning Beijing time, Junyang Lin, the technical lead of Qwen, posted that he was stepping down, with a goodbye to the team. By the afternoon Alibaba had called an emergency all-hands, and the reporting Simon Willison collected says the CEO attended. Binyuan Hui, who led the Qwen-Coder series, Bowen Yu, who led post-training research, and Kaixin Li, a core contributor on Qwen 3.5, VL and Coder, left the same day, along with a number of junior researchers.

The trigger, according to those reports, was a reorganisation in which a researcher from Google's Gemini team was promoted to lead Qwen. Lin's later message to the team was that they should continue as originally planned. We do not have more than that, and we are not going to speculate about the internal politics. What we can do is measure what was resting on the people who left.

Two weeks after Qwen 3.5

The timing makes the departure sharper. Qwen 3.5 landed on February 17 as a 397 billion parameter model weighing 807 gigabytes on disk. Over the following weeks the team shipped the smaller variants at 122, 35, 27, 9, 4, 2 and 0.8 billion parameters. The 2 billion model is 4.57 gigabytes in standard form and 1.27 gigabytes quantized. That ladder, from something that needs a rack to something that runs on a phone, is the reason Qwen became the default base for everyone else.

The people who left had just finished that release cycle. Whatever the next model is, it will be built by a team that lost its technical lead, its coding lead and its post-training lead in a single day. Post-training in particular is where most of a modern model's usable behaviour is set, and it is the part that is hardest to hand over because so much of it lives in the judgement of the people running it.

How much of the ecosystem is Qwen

Hugging Face's summer report gives the number we keep returning to. There are 151,448 repositories on the Hub derived from Qwen models, growing at roughly 180 to 210 new repositories a day through early 2026. That footprint is 2.6 times Meta's entire presence and 4.7 times the count of Llama repositories specifically. Downloads across all Qwen repositories total just over two billion.

The reason the derivative count matters more than the download count is who those derivatives are. They are the fine-tunes, quantizations, merges and distillations that individual developers, startups and academic labs build when they need a model they can modify and ship. The same report shows models under one billion parameters account for 83 percent of all-time downloads, and models over 100 billion for 1 percent. The small Qwen checkpoints are where the actual usage is, and the team that decided those sizes and licences just walked out.

The licence question is the part we would flag for anyone who builds on these weights. The report counts 178 Chinese releases above 20 billion parameters and finds 59 percent under Apache 2.0 and 22 percent under MIT. Qwen's permissive licensing was a policy choice by the team, and a new leadership can make a different one for the next release. Existing weights cannot be un-released, but the pipeline of future ones can be closed.

The single point of failure

We do not think this is a story about one company. The open-weight ecosystem has organised itself around a handful of Chinese labs shipping large models on permissive terms, and Hugging Face's own data shows Chinese labs' monthly parameter ceiling running between 754 billion and 2.78 trillion while US releases stayed under 130 billion in five of seven months. Within that group, Qwen was the one with the full size ladder and the widest derivative base. Concentration like that is efficient right up until the day it is not.

What worries us is less that Qwen 4 might be worse and more that the 150,000 downstream projects have no plan for a base model that changes character. A fine-tuning recipe that works on Qwen 3.5 may not transfer to a successor built by a different team with different post-training choices. Nobody has to port anything today. The question is how many of those projects would survive a year without new Qwen checkpoints, and we suspect the honest answer is most of them would freeze on the last release.

What we would want to see next

The constructive response is to reduce the concentration rather than to argue about it. Hugging Face's report shows Moonshot and MiniMax are already carrying most of the large-model download volume for their size class, and there are several labs with the compute to produce a full size ladder if they chose to. A second family with a 0.8 to 400 billion range on Apache terms would remove most of the risk.

The other thing worth doing is boring and important. If your product depends on a Qwen checkpoint, mirror the weights and the tokenizer and the licence file now, pin the exact revision, and write down the fine-tuning recipe in enough detail that someone could rerun it on a different base. The departures may turn out to be a blip and the next release may be fine. The point of the exercise is that you do not get to find out in advance.