Alibaba Unveils New AI Chip, Plans to Train New 5-10 Trillion Parameter Models

Alibaba plans to train an AI model with 5-10 trillion parameters and has unveiled a new AI accelerator designed to meet the computing demands of models at that scale. 

The announcements were made by Alibaba CEO Eddie Wu at the company’s Apsara Conference. 

The planned model is expected to be part of the Qwen family. Alibaba said Qwen 4 is currently in training, while the subsequent Qwen 4.5 and Qwen 5 series are projected to scale to between 5 trillion and 10 trillion parameters. That would be roughly two to four times the size of Alibaba’s current flagship Qwen 3.8 Max, which has 2.4 trillion parameters.

Wu said the Qwen team is working on model architecture and data optimisation to enable “more complex, longer-horizon tasks”. 

He also said the company has made progress on recursive self-improvement, in which models identify limitations, design experiments, and generate data to improve subsequent iterations. 

Alongside the model roadmap, Alibaba introduced the Zhenwu V900, a next-generation AI chip developed by its T-Head semiconductor division. 

Alibaba said the chip delivers three times the performance of the preceding M890 and can be deployed in clusters containing up to 500,000 chips for frontier-model training and inference. Mass production and commercial release are scheduled for the first quarter of 2027.

The V900 is part of Alibaba’s effort to build a domestic AI computing stack as Chinese companies face restrictions on access to advanced Nvidia processors. Alibaba previously introduced the M890 in May. 

Alibaba is also expanding the infrastructure intended to run these systems. Wu said Alibaba Cloud plans to increase its global data-centre capacity to more than 20 GW by 2032. The company said its M890-based AI supernodes can already perform inference on models exceeding 2 trillion parameters, a capability Wu said is currently available at only “a handful of companies globally”.

Earlier, Alibaba demonstrated Qwen3.7-Max running on the M890 for 35 hours and more than 1,000 tool calls, while subsequent M890 cloud deployments have been used to run models exceeding 2 trillion parameters, including Qwen3.8-Max and Kimi K3.