Alibaba unveils full-stack AI push with new Qwen chips

Alibaba's strategy focuses on meeting the rising demand for large-scale AI model training and inference by developing its own hardware and models. The Zhenwu V900 chip, designed by Alibaba's T-Head unit, aims to improve performance for these workloads, offering three times the capability of its predecessor and supporting multiple data precisions. This integration of chip design with cloud services reflects a broader industry trend of packaging AI systems as comprehensive stacks.

The company is also advancing its Qwen model family, with Qwen 4 currently in training and subsequent versions projected to reach trillions of parameters. A notable development is the progress in recursive self-improvement (RSI), demonstrated by Qwen3.8-Max's autonomous training optimization. This iterative approach to model development seeks to enhance performance through automated cycles of design, validation, and error diagnosis.

Beyond foundational models, Alibaba is expanding its offerings in speech, audio, and vision systems. New products like Qwen3.8-LiveTranslate aim to reduce latency in real-time interpretation, while Qwen-Audio-3.1-TTS-Next can generate both dialogue and ambient sound. These specialized tools suggest an effort to apply AI capabilities to specific industry needs, from content creation to translation services, within an 'agentic cloud' framework that organizes AI services across model, harness, and context layers.