Alibaba introduces new Qwen3.5 model for the “Agent AI Era”
Alibaba has introduced a new artificial intelligence model, Qwen 3.5, designed to independently perform complex tasks.
Alibaba has introduced a new artificial intelligence model, Qwen 3.5, designed to independently perform complex tasks.
Alibaba has introduced a new artificial intelligence model, Qwen 3.5, designed to independently perform complex tasks.
Alibaba has launched the Qwen3.5-397B-A17B, the first open-scale model in its Qwen3.5 series. Designed as a native visual-speech system, it has 397 billion parameters, although only 17 billion are activated in a single direct pass to improve efficiency.
The model uses a hybrid architecture that combines a sparse mix of experts with linear attention via Gated Delta Networks. The company says this design improves inference speed while maintaining high performance on reasoning, coding, and agent tests.
Multilingual support has expanded from 119 to 201 languages and dialects, thanks to a 250,000-token vocabulary and larger datasets for visual-text training. Alibaba claims the model delivers performance comparable to its much larger predecessors.

The cloud-based version of the model, Qwen3.5-Plus, is now available through Alibaba Cloud Model Studio . It has a 1 million token context window and built-in adaptive tooling. The Reinforcement Learning environment was scaled to prioritize the model’s ability to generalize tasks rather than narrow optimization for specific queries.
Infrastructure upgrades include an FP8 training pipeline and an asynchronous reinforcement learning platform to improve efficiency and stability. Alibaba positions Qwen3.5 as a foundation for multimodal agents supporting reasoning, search, and coding.



