Alibaba says next-gen Qwen is in training, targeting up to 10 trillion parameters
Alibaba disclosed a parameter-scale target for its next Qwen models at Yunqi; it is a roadmap statement, with real performance to be verified at release.
Original event 2026-09-22
At the Yunqi conference on September 22, Alibaba announced that Qwen4, built on a new architecture, is already in training, and that later versions such as Qwen4.5 and Qwen5 will scale to 5-10 trillion parameters.
The company also presented self-improvement results: Qwen3.8-Max reportedly ran 33 iterations with zero human involvement, raising its Artificial Analysis score from 40 to 45; the claimed 96% inference throughput gain and 42% chip area reduction are all vendor-reported figures.
The next-generation video generation model is due in November. A model in training has no verifiable results, and the parameter target may change; watch the actual release.