Alibaba has launched its latest AI accelerator, the Zhenwu V900, which significantly outperforms its predecessor, the M890, offering three times the performance. This chip, developed by Alibaba's T-Head division, is engineered for training and inference of large language models (LLMs) and features 216 GB of memory along with a high-speed inter-chip connection providing 1200 GB/s bandwidth. The company aims to begin mass production and commercial deployment of the Zhenwu V900 by the first quarter of 2027.
Accompanying the chip, Alibaba introduced a new server supernode utilizing the V900, capable of clustering up to 500,000 chips to support extensive AI model training and deployment. Currently, more than 650 clients from various sectors are leveraging Zhenwu-based solutions through Alibaba Cloud.
Future AI Models with Trillions of Parameters
Alibaba plans to integrate the new accelerators into its AI framework, with ongoing training on the Qwen 4 model and future versions, Qwen 4.5 and Qwen 5, expected to scale to between 5 and 10 trillion parameters. This is a significant increase from the current Qwen 3.8-Max model, which has 2.4 trillion parameters. CEO Eddie Wu emphasized that these advancements are aimed at enhancing the ability to manage complex tasks requiring long-term planning, a step towards achieving artificial superintelligence.
