Qwen3.5 debuts with hybrid architecture and expanded multimodal capabilities

Alibaba has released Qwen3.5-397B-A17B, the first open-weight model in its Qwen3.5 series. Designed as a native vision-language system, it contains 397 billion parameters, though only 17 billion are activated per forward pass to improve efficiency.

The model uses a hybrid architecture that combines sparse mixture-of-experts with linear attention via Gated Delta Networks. According to the company, this design improves inference speed while maintaining strong results across reasoning, coding, and agent benchmarks.

Multilingual coverage expands from 119 to 201 languages and dialects, supported by a 250k vocabulary and larger visual-text pretraining datasets. Alibaba says the model achieves performance comparable to significantly larger predecessors.

A hosted version, Qwen3.5-Plus, is available through Alibaba Cloud Model Studio, with a 1-million-token context window and built-in adaptive tool use. Reinforcement learning environments were scaled to prioritise generalisation across tasks rather than narrow optimisation.

Infrastructure upgrades include an FP8 training pipeline and an asynchronous reinforcement learning framework to improve efficiency and stability. Alibaba positions Qwen3.5 as a base for multimodal agents that support reasoning, search, and coding.

Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!