ai-acceleratorvendor
Zhenwu M890
T-Head's prior Zhenwu accelerator, with 144 GB of memory in Alibaba's May 2026 note.
Good forLarge-model training and inference on deployed Alibaba supernodes
Specifications
- Maker
- T-Head
- Region
- CHINA
- Class
- ai-accelerator
- Status
- shipping
- Claim level
- vendor
- Product
- Training and inference AI accelerator
- SKU
- M890
- Memory
- 144 GB
Alibaba’s May 2026 note rates the Zhenwu M890 at 144 GB of memory and 800 GB/s of inter-chip bandwidth, three times the Zhenwu 810E, with native formats from FP32 down to FP4. It does not state an FP16 peak, so none is recorded. TechNode’s account of the September Apsara Conference says M890 supernodes were already in large-scale commercial deployment and had run models above 2 trillion parameters, including Qwen3.8 and Kimi K3.
Name meaning:Zhenwu (真武) reads approximately as True Warrior or True Martial.
Context:The 144 GB figure is Alibaba's May 2026 claim. Support from FP32 down to FP4 is not an FP16 peak.