Alibaba introduced Qwen3.5 as an open-weight, natively multimodal model family. Its launch materials describe a 397B-parameter mixture-of-experts model with 17B activated parameters per forward pass, alongside reasoning, coding, agent and multimodal claims.
The release matters because it frames efficiency and multimodal agent work as one system-design question. That is different from treating a vision model as a bolt-on feature, but the practical trade-offs still depend on the exact model size, serving stack and workload.
Provider-reported benchmarks are useful context, not an independent league table. The official release and repository should remain linked whenever a specification or availability claim is discussed.