Alibaba is now playing the full stack game. At Apsara on Sept 22 the company confirmed Qwen 4 is in training and said Qwen 4.5 and Qwen 5 will scale to 5T to 10T parameters, up from 2.4T on today's Qwen3.8-Max. Existing Zhenwu chips already serve more than 650 customers, and Alibaba claims V900 clusters can scale to 500,000 cards. New cloud regions are coming to Türkiye, Finland and the Netherlands plus more capacity in Malaysia, Germany, the UAE, France and Hong Kong. Keep in mind every performance number is Alibaba's own claim, the 3x figure is measured against its own M890.

Alibaba's 😵 Minimax H3 Fun ControlNet Union (V2.0) Supports -Canny edge,Depth,Monocular ,HED edge detection, MLSD, PoseDW, ScribbleFree-hand / sketch lines, LayoutBounding-box layout, Grayscale (luminance) video 👇

Alibaba has open-sourced Logics-Parsing-V3, a 0.8B VLM for structured long-document parsing. Instead of resetting at every page, it carries a compact structural state forward as it reads — preserving context across pages without putting the entire document into one context window. This lets it recover document hierarchy, merge content split across pages, and link visual elements to related text in a single end-to-end model. It also handles complex layouts, scientific formulas, and chemical notation. On MPDocBench-Parse, it scores 85.26, ranking #1# — 4.46 points ahead of PaddleOCR-VL-1.5

Alibaba is accelerating its data center expansion in Europe and the Middle East, propelling its 20-gigawatt global infrastructure buildout during China’s escalating competition in AI chips and models with the US

Alibaba is going big: 5–10 TRILLION parameters for future Qwen models, with artificial superintelligence as the goal. Qwen 4 is already training, with that massive scale planned for Qwen 4.5 and Qwen 5. Alongside it, @Alibaba_Qwen unveiled its new Zhenwu V900 AI chip and announced plans to exceed 20 GW of data center capacity by 2032. Perhaps the most interesting part: Alibaba says its models are making progress at improving themselves, identifying weaknesses, running experiments and generating training data with limited human involvement.