Alibaba Launches Full-Stack AI Factory With Qwen3.7-Max

Alibaba Cloud has introduced a new "AI factory" stack, featuring the flagship Qwen3.7-Max large language model, the Zhenwu M890 AI accelerator, and the Panjiu AL128 server system [1]. By simultaneously launching these hardware and software components, the company aims to provide a vertically integrated platform that allows its AI models to autonomously optimize the very chips they run on [1].

Key takeaways

• The Qwen3.7-Max model demonstrated the ability to autonomously write and optimize performance-critical software for the Zhenwu M890 chip over a 35-hour period [1].
• Independent assessments from Artificial Analysis rank Qwen3.7-Max as the highest-scoring Chinese model, though it remains behind top-tier Western models like GPT-5.5 [1].
• The Zhenwu M890 chip is designed as a domestic alternative to high-end Western processors, with T-Head reporting over 560,000 shipments of its chip series to date [1].
• Alibaba has committed 380 billion yuan ($53 billion) to AI and cloud infrastructure over three years to support its full-stack strategy [1].

Integrating Hardware and Software for AI Autonomy

The core of Alibaba’s announcement is the operational synergy between its new model and its custom silicon. During a demonstration, Qwen3.7-Max ran for 35 hours on the Zhenwu M890 platform, executing 1,158 tool calls to iteratively optimize the chip's software stack without access to existing documentation [1]. Alibaba reports this process resulted in a 10x speedup on the model's Extend Attention kernel [1]. While these figures are first-party results, they highlight the company's strategy of tuning models specifically to its own hardware architecture [1].

The Zhenwu M890 chip itself represents a significant update to Alibaba’s semiconductor roadmap, featuring 144 GB of HBM3 memory and improved inter-chip bandwidth [1]. Industry analysts note that while the M890 trails the raw performance of leading Western chips like the Nvidia H200, it serves as a viable, high-performance replacement for the Chinese market [1]. The chip is integrated into the Panjiu AL128 server, which links 128 accelerators to support large-scale AI workloads [1].

Why it matters

Alibaba’s move toward vertical integration signals a long-term shift in China’s AI sector, moving beyond a reactive stance toward U.S. export controls [1]. Even as licensing pathways for some advanced U.S. processors have opened, Alibaba is prioritizing its own domestic stack to ensure it can tune software and hardware in tandem [1].

The company’s roadmap includes future iterations of the Zhenwu chip, with the V900 and J900 models planned for 2027 and 2028, respectively [1]. As Alibaba prepares to potentially spin off its T-Head semiconductor subsidiary through an IPO, the success of this "AI factory" model will depend on whether its performance gains can be independently replicated and whether it can maintain its competitive edge against both domestic rivals like Huawei and global frontier models [1].