Qwen quietly won open-source AI — then half-closed the door

Let's get the numbers out of the way. Qwen is reportedly the most-downloaded open model family on earth — over half of all open-source AI downloads worldwide, approaching a billion total by April 2026, with 433 open models published for download (MySummit). Mostly Apache 2.0: free, commercial, fine-tunable, no royalties. That is how you become default infrastructure in three years. You don't sell the model; you become the substrate everyone else builds on.

Then April 2026 arrives and on the 20th Alibaba ships Qwen3.6-Max-Preview — its first closed-weight, API-only flagship in three years of releases. No Hugging Face download. No ModelScope. No self-hosting. API-only, through Alibaba Cloud Model Studio (and they shut the free tier of Qwen Code the same day). The smaller open models kept shipping — Qwen3.6-27B (dense, Apache 2.0) and 35B-A3B (MoE, Apache 2.0) both landed the same month. So the family is not closed; the frontier is. There is a difference, and it matters. But the direction is clear: give away the on-ramp, charge for the supercar.

The logic is not complicated. Training a frontier model costs hundreds of millions of dollars in compute. Open weights do not pay that bill. Every open-weight lab — from Mistral to Meta — eventually does this calculation, and the result is always the same: the open tier becomes a customer-acquisition funnel for the paid frontier. Qwen3.6-Max-Preview is Alibaba's version of that moment. It arrived earlier than some expected for a Chinese lab that had positioned itself as the global open-AI default, which is precisely why it is worth noting.

Read that table and the 'is Qwen still open-source?' question answers itself: mostly yes, with a very visible 'no' at the top.

Here is the thing the cynical read misses. Qwen3.6-Max-Preview being closed does not undo what Qwen built. The installed base — approaching a billion downloads, 433 open models on the hub, an ecosystem of fine-tuned variants across every industry — is durable. A developer who already trained a product on Qwen weights does not migrate because the flagship went proprietary. The moat is the sunk cost of adoption. Alibaba knows this, which is why the closed pivot was only ever possible after the ecosystem was locked in.

What you should watch is whether the open tier stays genuinely useful or slowly becomes the demo. Right now the open models are real — Qwen3.6-27B reportedly outperforming a 397B predecessor on agentic coding is not a trivial benchmark claim. But if the performance gap between open and closed tiers is deliberately widened over successive Qwen generations, the open tier becomes a loss-leader and the 'most-downloaded open AI' brand becomes a funnel, not a fact. That transition, if it happens, will be slow and quiet. Watch the gap, not the headline.

So don't mistake the closed flagship for a retreat. Qwen already won the thing that matters — mindshare, installed base, the derivative ecosystem. A paywall on the very top model is just it starting to collect. The question is whether that collection stays bounded or gradually swallows the whole product. History says it usually expands. Watch the next generation.