Alibaba has put a countdown clock on the first piece of Qwen4 — the weights open tonight at 23:00 in Beijing
A day after the model page appeared and was pulled, Qwen has confirmed the schedule: Qwen3.8-Flash-Next, a multimodal mixture-of-experts model built on the next-generation Qwen4 architecture, opens at 23:00 Beijing time on 26 August in a standard and an FP8 version, on Hugging Face and ModelScope. The team's own framing is that this is a preview so the community can prepare for the full Qwen4 family, not a finished flagship. The figures circulating in Chinese tech press — 125 billion parameters plus 51 billion in n-gram embeddings, roughly 6 billion active per token, a training run about a ninth the cost of Qwen3.7-Plus — come from leaks rather than from Qwen, and no benchmark results have been published.
Why it mattersAlibaba is the only large lab still putting frontier-class weights out openly on a schedule, so an architecture preview from Qwen tells every open-source builder what the next year of cheap self-hosted models will look like. It also finally puts a date on the Qwen4 family, which has been the biggest unknown on the Chinese side since the spring.
Read in the app — free, in 9 languages
Related stories
Show the robot one video of a ten-minute job, and it does the ten-minute job — no retraining, nothing typed in2026-08-26Alibaba has posted a page for a model it calls a preview of Qwen4, due 26 August — the specs went up, then came down
2026-08-25The 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
2026-08-24