Alibaba Cloud's Zhenwu M890 Super Node Officially Commercialized, Supporting 20-Trillion-Parameter Large Models
Nashnova编辑部
Alibaba Cloud has put its Zhenwu M890 supernode into commercial service at its Inner Mongolia data center — the first Chinese system to run a model exceeding 2 trillion parameters on a supernode architecture, marking a shift from single-chip competition to full-system integration.
What exactly is the M890?
The Zhenwu M890 is a supernode — an architecture that lashes dozens of AI chips together with ultra-fast interconnects so they act as one giant machine. It is deployed in Ulanqab, Inner Mongolia, and available to global customers via Alibaba Cloud.
The core hardware uses chips designed in-house by DAMO Academy and Alibaba's T-Head semiconductor unit, supporting low-precision formats such as FP8 and FP4 — computing with fewer data bits per number to trade precision for speed.
This means → Alibaba Cloud has moved its homegrown silicon from "runs in the lab" to "customers can buy it" — commercial launch is the hard dividing line.
What changed from the last generation?
The key upgrade is interconnect scale: Alibaba's proprietary ICN Switch 1.0 chip expands a single node from the previous 16 accelerators to 64, with 800 GB/s chip-to-chip bandwidth and 9 TB of memory.
In plain terms = the old system was 16 workers moving bricks; the new one is 64, passing bricks to each other far faster.
Measured results: inference performance in agentic workloads rises up to 1.5×; training tasks in autonomous driving and embodied AI reach 3× the performance of the prior Zhenwu 810E.
What does "2 trillion parameters" mean here?
The M890 is the first Chinese system to run a model exceeding 2 trillion parameters on a supernode architecture. It has been validated on Alibaba's own Qwen 3.8, which carries 2.4 trillion parameters.
This means → a 2.4-trillion-parameter model is far too large for any single chip; the weights must be sliced and spread across dozens of tightly connected accelerators running in parallel — the 64-card interconnect exists precisely for this.
This reflects a stage where model sizes have outgrown single-machine capacity. Whoever gets "multi-card coordination" right first will be positioned to absorb the next generation of compute demand.
Where is Alibaba Cloud's strategic focus shifting?
Alibaba Cloud first showed the M890 at WAIC in July as an architecture demo; commercial launch means the system has cleared end-to-end validation from showroom to data-center floor.
The deeper signal: Alibaba Cloud's focus is moving from single-chip benchmark competition to system-level integration of chip + network + memory + software.
In plain terms = the old race was "whose chip scores higher"; the new race is "who can assemble chips, interconnects, memory, and software into a faster, cheaper machine" — whether this approach delivers a measurable edge in inference cost and user experience is the key thing to watch next.
Content is for reference only, not financial advice.