Alibaba Cloud's Lingjun Zhenwu M890 supernode entered commercial service on 12 August at the company's Ulanqab site in Inner Mongolia — billed as the first architecture in China serving language models above 2 trillion parameters.
The engineering claim
The M890 expands the scale-up domain from 16 cards to 64 using an in-house interconnect chip, ICNSwitch 1.0, with 800 GB/s between cards and support for FP8 and FP4. Alibaba rates it for mixture-of-experts inference on models up to 10 trillion parameters and claims three times the training performance of the previous-generation Zhenwu 810E.
What is running on it
Kimi K3 and Qwen3.8-Max are already served from the system. Because it is sold as a cloud instance rather than as hardware, customers reach models of this size without building their own AI data centre — which is the commercial point of the launch.
Unveiled earlier, in service now
The M890 was shown at WAIC before this. The 12 August event is general availability, so the accurate framing is that it entered service rather than that it was announced.
The question Alibaba did not answer
The domestic component that is confirmed is the interconnect. Alibaba has not disclosed which accelerators sit behind ICNSwitch, so claims that this is an all-domestic stack are unsupported. Given that export controls are the entire subtext, that omission is the most informative detail in the announcement.
Where the numbers come from
Every performance figure here is vendor-stated with no third-party benchmark. State media coverage has attached a self-reliance framing and supportive analyst commentary; that is editorial positioning around a product launch, not independent verification of it.