On 3 August we wrote that Alibaba was showing Qwen3.8-Max and promising its weights the following week. The promise has been kept: on 8 August the giant goes up for download, and three days before it, a small 27 billion model under Apache 2.0.
- Qwen3.8-27B lands on 5 August under Apache 2.0, with 27.8 billion parameters.
- Qwen3.8-2.4T-A95B is from 8 August: 2446 billion parameters, 95 billion active.
- The permissive licence is on the small one. On the big one, read the terms before you use it.
The promise from the start of the month has been kept. The big model's weights are out, and sooner than a week.
Next to it stands another model that gets talked about less. The big one is for headlines and for providers with a hall, and the small one is for you.
Why I am watching the small one
Twenty-seven billion parameters fit on a machine an average company can buy and put in a room of its own, with no hall and no transformer.
Apache 2.0 on top means you can put it in a product, sell it, and ask nobody. That is the difference between a model you look at and a model you use.
The big one runs on different logic. Mixture of experts means you hold 2.4 trillion parameters, but 95 billion do the work on any given request. The bill per answer is like a far smaller model, while the quality comes from the whole warehouse.
Handsome engineering, which still does not change the fact that to stand it up you need a hall.
The small print
The licence on the big one is not Apache. It is marked as other, which means the terms get read before anything is built on top. This is exactly where people get it wrong, and then wonder where the letter came from.
If you are setting out to build on open weights, take the small one and read the licence of everything you touch. A permissive licence is not something you assume because the files are free to download.