Alibaba's Qwen 3.5 vision encoder design takes a different path from the mainstream. Traditional Vision Transformers like LLaVA, MiniGPT, etc., mainly rely on single-layer output features, then align with the language model via a projection layer. Qwen 3.5's DeepStack Vision Transformer uses multi-layer feature fusion — integrating features from multiple intermediate layers of the encoder, rather than just looking at the final layer's output. Meanwhile, it uses Conv3D to treat video as a third dimension for processing, achieving native temporal modeling rather than post-hoc frame-stitching. The core benefit of this design: fine-grained texture and global semantics are no longer opposed, both can be preserved. For video Q&A, temporal reasoning, and other tasks, the improvement is significant. More importantly, this vision encoder isn't an independent external module, but is directly integrated into the language model's multimodal chain, embodying a "native multimodal" design orientation — solving the fusion problem at the architecture level rather than via post-training alignment.