Alibaba confirms Qwen 4 in training, unveils 3x-faster Zhenwu V900 chip at Apsara

Alibaba used its Apsara conference to lay out a full-stack roadmap. On models, Qwen 4 is in training. Qwen 4.5 and Qwen 5 are planned at 5–10 trillion parameters, and the current Qwen3.8-Max sits at 2.4 trillion parameters with more than 33 automated self-improvement cycles. On silicon, T-Head's Zhenwu V900 carries 216GB of memory and claims three times the performance of the Zhenwu M890, with mass production in Q1 2027. On infrastructure, Alibaba reiterated its plan for 20 GW of data-center capacity by 2032.

Alibaba Cloud also launched three products. Smart Studio and Smart Video target content workflows, and Smart Fusion claims about 50% lower token costs. The new Qwen3.8-LiveTranslate model rounds out the lineup, and Qwen3.8-27B became available through Nebius.

The strategy mirrors DeepSeek and Huawei's push to reduce dependence on NVIDIA. China's major labs are building their own chips and software stacks, which makes export controls a weaker lever over time.

The scrutiny is significant. CBS-reported research found systemic geopolitical filtering in Qwen models, along with weight-editing techniques to neutralize it. Separately, a US government website was found using Qwen after the FBI accused Alibaba of copying from Anthropic. Enterprises reportedly running Qwen, including Airbnb and Uber, face questions about these findings. Watch whether US agencies restrict Qwen use and whether the V900 meets its production timeline.