Alibaba Apsara: Qwen 4 enters training, Zhenwu V900 chip unveiled, 5-10T-parameter models planned

Alibaba used Apsara to present a full-stack AI roadmap covering models, chips, cloud and routing. On models, Qwen 4 has entered training, and the company outlined a multi-tier plan in which Qwen 4.5 and Qwen 5 reach 5 to 10 trillion parameters. The current flagship, Qwen3.8-Max, has 2.4 trillion. Alibaba also showcased recursive self-improvement experiments: Qwen3.8-Max ran more than 10,000 automated EDA tool calls and cut a chip bus module's area by 42%.

On hardware, the Zhenwu V900 chip is claimed to triple V800 performance, with commercialization in Q1 2027. That is part of China's broader push to reduce dependence on NVIDIA amid export controls. Alibaba Cloud added regions including Türkiye, Finland, Germany and the UAE, and introduced Smart Fusion multi-model routing, which it says cuts token costs by 50%. Multimodal releases included Qwen3.8-LiveTranslate, which cuts interpretation latency from 2.8 to 2.3 seconds, and Qwen-Audio-3.1-TTS-Next.

The announcements reinforce the narrative of Chinese labs closing the gap. Bloomberg Intelligence now puts the US-China benchmark difference at 3% after DeepSeek V4.1 Flash. Alibaba's 20GW-by-2032 data center target signals ambitions on par with US hyperscalers.

The skepticism is about trust and timelines. A CBS-reported study found pro-Beijing bias and censorship in Qwen, which powers services at some US companies, and that reignited concerns about building on Chinese open models. Ten-trillion-parameter goals are announcements, not shipped products. Watch Qwen 4's actual release date and licensing.