September 22, 2026
The Next Qwen Models Will Be Bigger: Alibaba Plans Models with Up to 10 Trillion Parameters
On September 22, 2026, Alibaba announced plans to train Qwen 4.5 and Qwen 5 with 5–10 trillion parameters. The same day, the company unveiled the Zhenwu V900 AI accelerator, which it says is three times faster than the M890 and will become commercially available in the first quarter of 2027.
In May 2026, Alibaba released the Zhenwu M890. The company now promises the V900 will deliver three times the performance, 216 GB of GPU memory, and inter-chip bandwidth of up to 1 200 GB/s. The accelerator supports FP8 and FP4 for model training and inference.
Not a standalone chip. Alibaba is assembling the V900 into a server supernode with an ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller. This system is designed for clusters of up to 500 000 cards. T-Head chips already serve more than 650 customers in automotive, finance, LLM, and three other industries.
Qwen 4 is already being trained. The company plans Qwen 4.5 and Qwen 5 at a scale of up to 10 trillion parameters, or up to four times larger than the current flagship. In an automated monthly run, Qwen3.8-Max completed 33 cycles and raised its Artificial Analysis score from 40 to 45.
In a chip-design experiment, the model worked for more than 60 hours, made more than 10 000 calls to EDA tools, and reduced the chip bus module area by 42% with no claimed loss of performance.
By 2032, Alibaba Cloud plans to exceed 20 GW of managed data center capacity worldwide.
Source
