Qwen 4 is already training: Alibaba targets a 5–10 trillion-parameter model

That is 2–4 times more than the current Qwen 3.8 Max flagship with 2.4 trillion parameters.

Not tomorrow. Qwen 4 is currently being trained, while the Qwen 4.5 and Qwen 5 series are planned to grow to 5–10 trillion parameters.

For these models, Alibaba also unveiled its own Zhenwu V900 chip: 216 GB of memory and 1200 GB/s between chips, with three times the performance of May's Zhenwu M890. A single cluster based on it can support up to 500 000 cards, but the chip will only enter mass production in the first quarter of 2027.

You can try it now. Alibaba opened an early preview of the Qwen4 architecture on August 26: Qwen3.8-Flash-Next weights are available on Hugging Face, including an FP8 variant, with a native context window of 262 144 tokens expandable to a million, and support for Transformers, vLLM, and SGLang. The current flagship is available through the API as qwen3.8-max. A million input tokens cost $2, output tokens cost $6, and cached input costs $0.25.

The company had already promised the funding for these plans in February: 380 billion yuan, about $53 billion, for AI and cloud over three years, and Wu warned analysts that Alibaba would exceed the plan. By 2032, Alibaba Cloud's data-center capacity is expected to exceed 20 GW.

Mass production of the Zhenwu V900 is promised for the first quarter of 2027.