Alibaba shares rise on unveiling trillion-parameter AI model
Alibaba’s Hong Kong shares rose after it released Qwen3‑Max‑Preview, a text‑only model with over one trillion parameters. The company says it outperforms Qwen3‑235B on internal tests and is available via Alibaba Cloud and OpenRouter.
Alibaba Group’s stock in Hong Kong (9988) gained as much as 4% to HK$138.0 on Monday (Sept. 8) after the company released Qwen3‑Max‑Preview, described as its largest and “most intelligent” AI model so far. The text‑only model exceeds one trillion parameters and, according to Alibaba, outperforms its previous flagship Qwen3‑235B on internal tests.
The preview release arrives as Alibaba steps up competition with OpenAI and Google DeepMind, joining the cohort of trillion‑parameter systems. The firm published scores suggesting Qwen3‑Max‑Preview topped Moonshot AI’s Kimi K2, a non‑reasoning version of Anthropic’s Claude Opus 4, and DeepSeek V3.1 on five benchmarks; Alibaba did not publish an official technical report alongside the preview.
Scaling works - and the official release will surprise you even more,the company said.
Qwen3‑Max‑Preview is available through Alibaba Cloud and the OpenRouter marketplace. The model is not open‑sourced at this stage. Tiered API pricing on Alibaba Cloud starts at US$0.861 per million input tokens and US$3.441 per million output tokens, making it one of the pricier Qwen models. For reference, Qwen3‑235B‑A22B‑2507 is listed at US$0.287 (input) and US$1.147 (output) for its non‑thinking version.
On social media, Alibaba AI engineer Binyuan Hui said a “thinking” version is on the way.
Alibaba has committed ¥380 billion (≈ US$52B) to AI infrastructure over the next three years and says AI‑related products have posted triple‑digit growth for eight consecutive quarters. Reports last week said the firm is developing a home‑grown AI processor to reduce reliance on Nvidia chips, amid tighter scrutiny of foreign semiconductors. In its fiscal June‑quarter (reported late August), cloud revenue beat expectations, though overall revenue was pressured by tougher conditions in core e‑commerce.
Near‑term markers include an official technical report for Qwen3‑Max‑Preview and independent benchmarks, the launch of the “thinking” variant, signs of enterprise adoption and workload mix, and disclosures on the in‑house AI processor program and its availability.
The company also said Qwen3 models have surpassed 20 million downloads and 100,000 derivative models on Hugging Face, underscoring developer uptake for its open‑source releases even as the new model remains closed for now.
