Alibaba shows the Zhenwu M890 chip and Qwen3.7-Max
On 20 May 2026, at the Alibaba Cloud Summit, Alibaba announced the flagship model Qwen3.7-Max ('soon' through Model Studio), T-Head's Zhenwu M890 accelerator and the Panjiu AL128 server, which puts 128 accelerators in one rack and, the company says, is already available in Model Studio for the Chinese market.
Why it matters
A Chinese cloud company offers its own accelerator and its own model as one stack. An editorial assessment: every figure is Alibaba's own word; neither the chip nor the model has an outside measurement.
Figures the company gives. Zhenwu M890: three times the performance of the Zhenwu 810E, 144 GB of memory, 800 GB/s between chips, support from FP32 down to FP4; the ICN Switch 1.0 switching chip, 25.6 Tbps; more than 560,000 Zhenwu units delivered and more than 400 external customers. For Qwen3.7-Max: up to 35 hours of continuous operation and over 1,000 tool calls. What the record does not claim. None of these numbers was measured by a third party. The page does not say whether Qwen3.7-Max weights are published. The release text carries the date 'May 20, 2025', an evident slip; the page's date is 20 May 2026.