Alibaba Shares Big Model Updates: Qwen4 Is Training, RSI Takes Early Shape

At the 2026 Yunqi Conference, Alibaba shared a series of big model updates. Recursive self-improvement (RSI) has started to enter model training, inference, and chip-model co-design. Qwen4, built on a new architecture, is now in training. Future versions like Qwen4.5 and Qwen5 will scale to 5-10T parameters.

Qwen3.8-Max ran over a month of self-iteration with zero human involvement, completing 33 effective rounds. Its Artificial Analysis score rose from 40 to 45, putting it in the same top tier as Claude and GPT's strongest models. On inference optimization, it independently adapted the next-gen Qwen3.8-Flash inference framework on T-Head's new GPU, boosting single-instance throughput by 96%. In chip-model co-design, it ran autonomously for over 60 hours and cut physical implementation area by 42%.

Alibaba's next-gen video generation model will launch in November, moving toward longer, more controllable, and smarter output. The AI phone full-stack solution Qwen Intelligence was also released, optimized for mobile scenarios.