Alibaba confirms Qwen 4 is in training and unveils Zhenwu V900 chip at Apsara 2026
Alibaba used its Apsara 2026 conference to present itself as a full-stack AI company, covering models, chips, cloud capacity and tools. The headline model news is that Qwen 4 is now training. The roadmap behind it names Qwen 4.5 and Qwen 5, which Alibaba says will reach 5 to 10 trillion parameters. For scale, the current flagship Qwen3.8-Max sits at 2.4 trillion. Alibaba also launched Qwen3.8-LiveTranslate for simultaneous interpretation, plus Smart Studio, Smart Fusion and Smart Video tools on Alibaba Cloud.
On hardware, the T-Head division introduced the Zhenwu V900, a processor for both training and inference with 216GB of memory and native FP4 precision. FP4 matters because it is the low-precision format NVIDIA's Blackwell generation uses to boost inference throughput. Alibaba also showed what it calls recursive self-improvement milestones: Qwen3.8-Max ran 10,000 EDA operations during chip design work and cut bus footprints by 42%. In other words, the model is helping design the chips that will train its successors. On the cloud side, Alibaba targets more than 20GW of data center capacity by 2032 and is opening regions in Türkiye, Finland and the Netherlands.
The competitive picture is a race toward sovereign Chinese compute. In the same week, DeepSeek and Huawei open-sourced Ascend tooling meant to replace CUDA. Alibaba is betting on its own silicon instead of Huawei's, but both moves point the same way: less dependence on NVIDIA. Meanwhile, Qwen's open weights keep spreading in the West. A practitioner this week ran the 125B Qwen3.8-Flash-Next MoE on a single RTX 3090 Ti, and Ethan Mollick noted that Europe's new sovereign model was partly fine-tuned on Qwen- and GLM-generated data.
There are caveats. Parameter targets are roadmap promises, not shipped models, and Alibaba has given no Qwen 4 release date. A study reported this week also found pro-Beijing bias and censored answers in Qwen, which could slow adoption among US companies building on Chinese models. Watch for independent Zhenwu V900 benchmarks and whether Qwen 4 stays open-weight.