Alibaba has posted a page for a model it calls a preview of Qwen4, due 26 August — the specs went up, then came down
A repository named Qwen/Qwen3.8-Flash-Next appeared on Hugging Face and on ModelScope, Alibaba's own model hubs, describing the model as a preview of the Qwen4 architecture with an expected release of 26 August. Developers who read the ModelScope listing before it was edited reported a mixture-of-experts design at 125 billion total parameters with about 6 billion active per token, roughly 51 billion parameters of n-gram embeddings, and a new sparse attention scheme; those figures were pulled from the page soon after and Alibaba has not confirmed them. The stated intent is to hand the architecture to developers ahead of the full Qwen4 family rather than to ship a flagship.
Why it mattersAlibaba is the largest publisher of open-weight models in the world, and shipping the architecture before the flagship is not a scheduling accident. Inference engines, quantisation tools and fine-tuning stacks each need weeks of work to support a new attention scheme, so releasing a small preview first means the open ecosystem is ready on the day the real models arrive. It also hints at where Qwen4 is going: if the removed figures are right, around 6 billion parameters active per token is a design aimed at running fast on modest hardware rather than at winning a parameter count.
Read in the app — free, in 9 languages
Related stories
The final training run cost about $450,000. The company that owns Westlaw now has its own frontier model.2026-08-25A video model you steer with the arrow keys while it is still generating — and the company admits it looks worse than the offline ones
2026-08-25Alibaba's new video model takes your slide deck as the input and hands back thirty seconds of film
2026-08-24Alibaba trained a model on a rack of 100 real phones — it now finishes 92 out of 100 tasks on an actual handset
2026-08-23An eight-billion-parameter model that reads pictures and draws them, at 4K, with an Apache licence — and the community had it running in two days
2026-08-22