Qwen3.5

Qwen3.5 is a multimodal mixture-of-experts model with 397 billion total parameters and 17 billion active parameters, built on a hybrid Gated Delta Network architecture that enables long-context inference with near-linear compute scaling. The flagship of the Qwen3.5 family, it delivers state-of-the-art performance across knowledge, reasoning, coding, vision-language understanding, and agentic tasks — with native multimodal support and reasoning mode built in. Served as 397B-A17B and 35B-A3B.