HANGZHOU — Alibaba Cloud today unveiled Qwen3-Max, its largest large language model to date with over a trillion parameters and 36 trillion training tokens, at the annual Apsara Conference. CEO Eddie Wu framed the release as part of a full-stack ‘AI + cloud’ strategy, reiterating its RMB 380 billion (about $52.9 billion) three-year AI and cloud infrastructure plan announced in February. The company also deepened its partnership with Nvidia to power so-called ‘Physical AI.’

Qwen3-Max offers both instruct and thinking modes, scoring 69.6 on the SWE-Bench coding benchmark—on par with leading closed-source models—and showing strong results on the Tau2-Bench agent evaluation. Alibaba released several other models spanning visual, multimodal, and speech families (Qwen3-VL, Qwen3-Omni, Qwen3-Coder, Wan2.5 preview, and Fun). Wu declared that large models are ‘the operating system of the AI era’ and that Alibaba Cloud aims to build a ‘super AI cloud’ as the next computer, with an artificial superintelligence roadmap spanning emergent intelligence, autonomous action, and self-iteration.

Alibaba shares jumped over 9% in Hong Kong trading following the announcements. AI-related income has grown at triple-digit rates for eight consecutive quarters, and Alibaba Cloud holds 35.8% of China’s AI cloud market, according to Omdia. The splashy releases come as competition heats up with Tencent, ByteDance, and DeepSeek, and as global AI capex surges—Nvidia just announced a $100 billion investment into OpenAI days before this conference. Wu predicted global AI investment will exceed $4 trillion over five years, and warned that data center energy consumption at Alibaba Cloud will grow tenfold by 2032.